Tangled Up in Blue: Evaluating Milei’s Macroeconomic Reforms in Argentina Current Effects and Future Outlook Fredrik Eneroth & Benjamin Larsson Abstract: In December 2023, Argentina entered a new economic chapter under President Javier Milei, who initiated sweeping reforms to combat hyperinflation, fiscal deficits and rising poverty. This thesis evaluates the short- term macroeconomic effects of Milei’s reform agenda and assesses Argentina’s likely trajectory ahead of the pivotal 2025 midterm elections. Using a structural Vector Autoregression (VAR) model, the analysis isolates the effects of key policy interventions, including fiscal austerity, monetary tightening, deregulation and partial currency liberalization on inflation, unemployment, real wages, poverty and GDP. Counterfactual simulations estimate how these indicators would have evolved absent Milei’s reforms, providing a causal interpretation of observed macroeconomic changes in 2024. The results suggest that the reform package contributed to a sharp decrease in inflation and a primary fiscal surplus, but with adverse short-run effects on real wages and poverty. The effects on employment were more muted, with mixed signals across quarters. The VAR model is also used to forecast macroeconomic conditions into late 2025, offering insights into voter sentiment ahead of the elections. While the government has prioritized stability to sustain early gains, the outlook remains fragile. If real incomes and growth fail to recover meaningfully, public support may erode, constraining future reforms. This thesis contributes to the literature on political economy and shock therapy by highlighting the tradeoffs between rapid liberalization and political sustainability in emerging democracies. Bachelor’s thesis in Economics, 15 credits Spring Semester 2025 Supervisor: Xiyu Jiao Department of Economics School of Business, Economics and Law University of Gothenburg Table of Contents TABLE OF CONTENTS .................................................................................................................................... 2 1. INTRODUCTION ................................................................................................................................... 4 1.1 MILEI’S REFORM AGENDA .......................................................................................................................... 4 1.2 2025 MIDTERM ELECTIONS ....................................................................................................................... 5 1.3 AIM AND RESEARCH QUESTIONS ................................................................................................................. 5 2. LITERATURE REVIEW ........................................................................................................................... 6 3. METHODOLOGY .................................................................................................................................. 8 3.1 DATA AND VARIABLES ............................................................................................................................... 8 3.2 DATA FREQUENCY HARMONIZATION .......................................................................................................... 11 3.3 LOGARITHMIC TRANSFORMATIONS AND DIFFERENCING .................................................................................. 12 3.4 MODEL AND PARAMETER SELECTIONS ........................................................................................................ 12 3.4.1 Choice between VAR and VECM .................................................................................................... 13 3.4.2 Lag Order Selection ....................................................................................................................... 14 3.5 LEVEL RECONSTRUCTION FOR INTERPRETABILITY ........................................................................................... 15 3.6 ISOLATION OF POLICY EFFECTS .................................................................................................................. 16 3.7 MODEL VALIDATION ............................................................................................................................... 17 3.7.1 Structural Break Diagnostics ......................................................................................................... 17 3.7.2 Search for Heteroskedasticity ....................................................................................................... 18 3.7.3 LM-testing ..................................................................................................................................... 20 3.7.4 Eigenvalue testing ......................................................................................................................... 21 4. RESULTS ............................................................................................................................................ 22 4.1 MODEL AND PARAMETER SELECTIONS ........................................................................................................ 22 4.1.1 Stationarity Testing ....................................................................................................................... 22 4.1.2 Lag Order ...................................................................................................................................... 24 4.2 POLICY EFFECTS DURING 2024 ................................................................................................................. 25 4.2.1 Counterfactual Scenario................................................................................................................ 26 4.2.2 Policy Effects Month by Month ..................................................................................................... 29 4.2.3 Total Policy Effects ........................................................................................................................ 30 4.3 FORECAST FOR 2025 .............................................................................................................................. 31 4.4 MODEL VALIDATION ............................................................................................................................... 37 4.4.1 Structural Break Diagnostics ......................................................................................................... 37 4.4.2 Search for Heteroskedasticity ....................................................................................................... 39 4.4.3 LM-testing ..................................................................................................................................... 43 2 4.4.4 Eigenvalue testing ......................................................................................................................... 44 5. DISCUSSION AND CONCLUSIONS ....................................................................................................... 46 5.1 INTERPRETATION OF RESULTS .................................................................................................................... 46 5.2 MODEL UNCERTAINTY AND ROBUSTNESS .................................................................................................... 47 5.3 POLITICAL AND ECONOMIC CONTEXTUALIZATION .......................................................................................... 48 5.4 POLICY IMPLICATIONS .............................................................................................................................. 49 5.5 SUGGESTIONS FOR FUTURE RESEARCH ........................................................................................................ 50 REFERENCES ............................................................................................................................................... 52 3 1. Introduction In December 2023, Argentina’s economy was in a state of crisis, marked by runaway inflation, fiscal imbalances and social distress. Consumer prices had surged by over 211% year on year, the nation’s highest inflation rate for decades (Paqué & Holtzmann, 2024; Falcone, 2024). The public finances were similarly dire: the overall fiscal deficit reached an estimated 13% of GDP in 2023 (Harguindeguy, 2024). Wages failed to keep pace with prices, contributing to a sharp rise in poverty to around 45% of the population by late 2023 (Paqué & Holtzmann, 2024). Amid fears of hyperinflation and a deep recession, many voters turned to Javier Milei, a libertarian economist and political outsider, who campaigned on a promise of radical reforms to break Argentina’s cycle of economic crises and restore stability. 1.1 Milei’s Reform Agenda Once sworn in, Milei quickly launched sweeping reforms to stabilize the economy. A cornerstone among the reforms was fiscal tightening. His administration slashed cut subsidies on energy, transportation and other services, downsized the bureaucracy and even shuttered several ministries (Green, 2024). These and other austerity measures were socially painful but yielded rapid results. The government achieved a primary budget surplus within the first quarter of 2024 (Falcone, 2024). Milei also moved to end the monetary financing of deficits by the central bank, halting the money-printing presses that had fueled inflation (Harguindeguy, 2024). Beyond these reforms aimed at stabilization, Milei advanced structural reforms through an omnibus legislative package known in English as the “Law of Bases and Starting Points for the Freedom of Argentines”. The bill, just like many other proposals, faced political and institutional constraints. Once approved in mid-2024 in a trimmed down form it still encompassed hundreds of measures to liberalize the economy (Paqué & Holtzmann, 2024). For instance, the law relaxed labor regulations, extending probationary periods for new hires and easing severance rules to encourage employment (Soltys, 2024). It also lifted price controls, reduced red tape and authorized privatization of numerous state owned enterprises, including the national airline and energy companies (Green, 2024). In the financial realm, Milei took steps toward currency liberalization, moving to unify exchange rates and loosen capital controls, with the ultimate goal of dollarization (Associated Press, 2024). 4 1.2 2025 Midterm Elections The October 2025 mid-term elections will be a pivotal juncture for Milei’s reform agenda. With half of the lower house and one third of the Senate up for renewal, the midterms will determine whether Milei can convert his personal mandate into a stronger legislative bloc with a working majority (Werner, 2025). Such a shift in the balance of power would enable the government to pursue deeper structural reforms that have so far been politically unattainable, including comprehensive tax simplification and dollarization. Conversely, a poor showing could leave Milei without sufficient votes in Congress to pass further major reforms. Considering these stakes, Milei’s economic team has calibrated the reform timetable to maintain short-term stability ahead of the elections. In particular, the government has held off on completely lifting foreign exchange controls, wary that a devaluation could spike inflation (Werner, 2025). 1.3 Aim and Research Questions Given the history of recent events affecting and the upcoming election in Argentina, this thesis aims to empirically assess both the effectiveness and the likely trajectory of Javier Milei’s economic reforms. In this context, the thesis provides a timely empirical contribution to the broader debate on the political sustainability of economic shock therapy in fragile democracies. The central research question is twofold: (1) How did Milei’s reforms influence Argentina’s key macroeconomic indicators during his first year in office? (2) What is the likely macroeconomic outlook in the run-up to the 2025 midterm elections? To address these questions, the study employs a Vector Autoregression (VAR) model to create a counterfactual simulation for 2024 based on previous trends, isolating the effects of Milei’s reforms. The VAR model is also used to forecast Argentina’s macroeconomic conditions into late 2025, closer to the midterm elections. 5 2. Literature Review Forecasting macroeconomic indicators in emerging economies is complex and has attracted significant research attention. Aguiar and Gopinath (2007) describe emerging market business cycles as being largely driven by permanent shocks, unlike the transitory ones in advanced economies, meaning that forecasting is difficult because patterns can shift abruptly. Procyclical fiscal and monetary policies also contribute to forecasting errors. These policy swings are often unanticipated by standard models. The original theory by Nordhaus (1975) proposes that incumbents would utilize expansionist fiscal or monetary policy ahead of elections to boost the economy, leading to pre-election growth and post-election inflation. Brender and Drazen (2005) identify that new democracies have a prevalent tendency towards political budget cycles, where electorates are less inclined to sanction pre-election fiscal irresponsibility. Recently, Shi and Svensson (2006) with a sample of 85 countries found election-year deficits and outlays peaks were typical among developing nations, primarily due to weaker institutional restraints along with more information asymmetry. This holds particular relevance for Latin America, where re-democratization took place in the 1980s and the 1990s. Ames (1987) reported a significant rise in government spending by about 6% during election years in Latin America. In nations such as Argentina, where prior to elections expansionary changes are common, neglect of the political cycle can lead to systematic forecast errors such as mispredicting post- election inflation peaks. Forecasts can be made more accurate by allowing for the use of election dummies or by controlling for regime-switching behavior in models. Cermeño, Grier and Grier (2010), for example, discovered governments in Latin America tended to peg or overvalue exchange rates prior to elections, then depreciate. Dˇrazen (2000) further indicates that forecasters in these environments tend to condition assumptions for government spending or inflation on when elections are held. Frankel (2017) suggests incorporating measures like policy uncertainty indices or commodity price forecasts to improve predictive accuracy, given their strong influence in emerging markets. 6 Even when policy improvements are seen, forecasting in emerging markets remains difficult due to continued volatility and weaker data quality. Shorter, noisier time series and frequent data revisions lead to less precise forecasts. Ho and Mauro (2016) point out that international institutions have often been overly optimistic, underestimating downside risks and structural challenges. Standard models often fall short in these high-volatility environments. To better capture uncertainty, forecasters have turned to models like GARCH and stochastic volatility frameworks that account for variable risk. Mandalinci (2017) finds that while no model is consistently superior, those with time-varying parameters or stochastic volatility perform better in volatile environments. Vector Autoregression (VAR) models (Sims, 1980) are widely used due to their ability to model interdependencies without strong assumptions. While tools like VAR, ARIMA and structural models are commonly used, emerging market forecasting requires adjustments to account for volatility and regime shifts. Techniques like accounting for structural breaks, using Bayesian shrinkage in small sample VARs or averaging across models can improve performance. Forecasters also often complement models with expert judgment about commodity prices, political risk and other regional dynamics. High-inflation settings test traditional models like the Phillips curve. Generally, inflation remains hard to forecast and simple models often rival more complex ones (Duncan and Martínez-García, 2018). Forecasting unemployment typically involves structural relationships, e.g. Okun’s law or VAR models incorporating inflation and output. Forecasting real wages and poverty is less developed but important. Real wage growth is often inferred from inflation and productivity (Mihaljek & Saxena, 2010), while poverty forecasting uses macro-micro simulations combining GDP, inflation and employment projections (World Bank, 2020). 7 3. Methodology 3.1 Data and Variables The data used in this study was collected from Argentina’s national statistics bureau, the Instituto Nacional de Estadística y Censos (INDEC). INDEC is the primary source for official economic and social statistics in Argentina, widely utilized by researchers, policymakers and international organizations. Its credibility and comprehensive coverage make it a reliable basis for analyzing macroeconomic indicators. The five indicators, inflation, wages, poverty, unemployment and GDP, capture distinct but interrelated dimensions of Argentina’s short-run macroeconomic adjustment process. Inflation and wages jointly reflect price and income dynamics, which are particularly important in the context of high inflation and real income volatility. Poverty and unemployment are key indicators of social welfare and labor market performance, both of which are expected to be sensitive to structural reforms. Finally, GDP represents overall economic output and provides a broad measure of aggregate performance. This particular set of variables reflects a deliberate trade-off between economic completeness and model tractability. While a broader system incorporating monetary and financial variables, such as exchange rates, interest rates and fiscal balance measures, could potentially enrich the model, their inclusion would substantially increase the dimensionality of the VAR system. Given that the number of parameters that need to be estimated in a VAR grows quadratically (𝑝𝑘2parameters, for 𝑘 variables and lag order 𝑝), such additions would pose serious risks of overfitting and multicollinearity, particularly in a sample with limited time observations. For example, expanding the model from 5 to 7 variables with a single lag would increase the total number of parameters from 25 to 49, significantly reducing the degrees of freedom available for reliable estimation and hypothesis testing. Several other macroeconomic indicators, e.g. exchange and interest rates, were deliberately excluded. While these variables are important in broader macroeconomic frameworks, they tend to be highly responsive to exogenous global shocks. Including such series would not have been ideal given the method this thesis applies, i.e. creating domestic forecasts based on 8 earlier data, since global variables are difficult to model. Imbedding such variables in the model would have risked both large errors concerning them and downstream other variables. The time series dataset consisting of five variables covers the period from January 2017 to as far as March 2025. Monthly data are used for the Consumer Price Index (CPI) and the nominal wage index, quarterly data for the unemployment rate and GDP, and semi-annual data for the poverty rate. As of April 2024, the CPI data are available through March 2025, the nominal wage index through January 2025 and poverty, GDP and unemployment data through December 2024. Consequently, the number of observations for the variables differ, as shown in table 1. Table 1 further shows some distributional properties of the variables during their respective periods. CPI ranged from 102 to 8353, adjusted GDP from 564 to 738 billion pesos (2004 prices) and Nominal Wage Index from 104 to 5934. The poverty rate spanned from a low of 22% to a high of 53%, with a mean of 36%, while the unemployment rate had a low of 6% and a high of 13%, with a mean of 8%. Table 1 – Summary statistics. CPI Poverty rate Unemployment Adjusted Nominal (December rate GDP (BN Wage Index 2016 = 100) pesos, 2004 (1 October prices) 2016 = 100) Observations 99 17 33 33 97 Mean 1513 36 8 699 1026 Standard deviation 2322 7 2 36 1506 Min 102 22 6 564 104 25% 193 32 7 691 175 50% 416 37 8 704 329 75% 1332 41 10 722 973 Max 8353 53 13 738 5934 Consumer Price Index (CPI) 9 The Consumer Price Index (CPI) captures monthly changes in the overall price level of goods and services purchased by households across Argentina. The series used in this study corresponds to the national level index. It is standardized with value 100 in December 2016. Poverty Rate The poverty rate refers to the percentage of individuals living below the official poverty line, defined by INDEC as the income required to afford a basket of goods and services essential for basic living. It consists of data on individuals rather than households. Unemployment Rate Unemployment is measured as the share of the economically active population (>14 years old) that is not employed but is actively seeking work. It excludes those outside the labor force, such as full-time students, retirees and homemakers. The indicator corresponds to Desocupación abierta as reported by INDEC through the Encuesta Permanente de Hogares. Gross Domestic Product (GDP) Gross Domestic Product is measured in real terms, adjusted for inflation using 2004 prices (pesos), to distinguish changes in the volume of economic activity from price changes. It is seasonally adjusted and reported by the Ministry of Economy in millions of pesos under the heading PIB – Oferta y demanda globales (series desestacionalizadas). Nominal Wage Index The nominal wage index reflects the evolution of average gross salaries across the Argentine economy. The index is standardized to 100 on October 1, 2016, and includes both formal and informal sectors. In this study, it is utilized alongside the CPI to construct real wages. Construction of Real Wage Index As a step in variable selection, the Nominal Wage Index (𝑊) was exchanged for a Real Wage Index (𝑊𝑟). The Real Wage Index was considered more important due to its direct relevance to questions of purchasing power, labor market outcomes and household welfare. The real wages were constructed from the Nominal Wage Index and the CPI as follows: 10 𝑊𝑟 𝑡𝑊𝑡 = × 100 (1) 𝐶𝑃𝐼𝑡 Equation (1) yields a Real Wage Index scaled for interpretability. The multiplication by 100 is a scalar adjustment that preserves the magnitude and comparability of the resulting series to that of the Nominal Wage Index. 3.2 Data Frequency Harmonization A key step in any time-series econometric analysis involves preparing the data in a format compatible with the structure and assumptions of the intended model. In the case of this thesis, the empirical analysis is built around a Vector Autoregression (VAR) model, which requires all included variables to be observed at the same frequency and aligned on a monthly time scale. However, the dataset comprises macroeconomic indicators with heterogeneous reporting frequencies: the Consumer Price Index (CPI) and the wage index are reported monthly, while the unemployment rate is observed quarterly and poverty rates are only available semi-annually. This frequency mismatch poses a significant obstacle to estimation, as VAR models assume synchronous updates across time for all endogenous variables. A prerequisite for VAR estimation is that all variables are aligned in time and reported at the same frequency. In this study, several macroeconomic indicators differ in temporal resolution: CPI and wages are monthly, unemployment is quarterly and poverty is semi- annual. To harmonize these series, linear interpolation was applied to the unemployment and poverty variables, producing monthly estimates that preserve the gradual evolution typical of macroeconomic indicators. The interpolation method assumes a constant rate of change between observations, formally expressed as the following: 𝑥1 − 𝑥0𝑥𝑡 = 𝑥0 + × (𝑡 − 𝑡0), 𝑡 ∈ (𝑡0, 𝑡1) (2) 𝑡1 − 𝑡0 Observed values were assigned to the final month of each reporting period, e.g. March, June, September, and December for quarterly data. This decision reflects standard practices by statistical agencies and supports temporal alignment within the VAR framework. While necessary, interpolation can introduce bias by smoothing short-term volatility, particularly during periods of economic turbulence. This is a relevant concern in the 11 Argentine context and results involving interpolated series have been interpreted cautiously. Additionally, since some series were not reported beyond 2024, data from 2025 was excluded to maintain consistency and avoid distortions in model estimation and forecasting. 3.3 Logarithmic Transformations and Differencing Variables exhibiting trend-like or stochastic exponential growth, such as the Consumer Price Index (CPI), real wages and gross domestic product (GDP), were first transformed using natural logarithms. This transformation helps linearize exponential trends and stabilize variance, making the data more suitable for most time-series modeling, e.g. VAR. Modeling in differences ensures stationarity and addresses the risk of spurious regression that arises from estimating relationships among non-stationary series in levels. First differencing was then applied to the logged series, resulting in a form that approximates monthly percentage changes: 𝑦𝑡 − 𝑦𝑡−1Δ ln(𝑦𝑡) ≈ (3) 𝑦𝑡−1 Equation 3 could for example be applied for inflation in the following way: inflation = 𝐶𝑃𝐼 −𝐶𝑃𝐼 𝑡 𝑡−1 ≈ Δ ln(𝐶𝑃𝐼𝑡). Thus, through first difference logs, CPI, GDP and Real Wage 𝐶𝑃𝐼𝑡−1 Index can be interpreted as month-to-month percent changes. In contrast, poverty and unemployment rates, expressed as bounded rates between 0 and 100, were differenced in levels without logarithmic transformation. These variables do not exhibit multiplicative behavior and absolute changes in percentage points are more meaningful in both economic and policy contexts. 3.4 Model and Parameter Selections Given the difficult task of estimating the effect of many radical reforms in a volatile economic environment, it was deemed more feasible to simply forecast how the Argentine Economy would develop based on historic trends. This still allows the thesis to answer the two research questions thanks to the data available. Using pre-treatment data, one can forecast a counterfactual economic progression during 2024, allowing an answer to the first research question. Using post-treatment data, i.e. the data from 2024, one can forecast the 12 economic progression during 2025 given no further reforms, answering the second research question. The modeling strategy in this study focuses on structural time series analysis rather than machine learning, due to the nature of the research questions. While advanced forecasting techniques such as Long Short-Term Memory (LSTM) networks can capture complex nonlinear patterns, they are often treated as black-box models and provide limited insight into the structural relationships among economic variables. Moreover, they typically require large datasets and stable regimes to perform reliably, conditions that are not met for Argentina. Given the objective of estimating policy effects and generating interpretable counterfactuals, the analysis centers on classical econometric frameworks. The appropriate choice is thus likely some form of Vector Autoregressive (VAR) model, given their prevalence. With a multivariate time series 𝒀𝑡, the standard Vector Autoregression (VAR) in first-differences model can be expressed as the following: 𝑝 ∆𝒀𝑡 = (∑ 𝜱𝑖∆𝒀𝑡−𝑖) + 𝒖𝑡 (4) 𝑖=1 In equation (4), 𝑝 denotes the number of lags, ∆ the first-difference operator, 𝜱𝑖 coefficient matrices and 𝒖𝑡 a white noise term. This foundational VAR model can, if deemed beneficial, be adjusted to for example imbed long-run trends through levels (Vector Error Correction Model). The number of lags (𝑝) is an important parameter choice either way. 3.4.1 Choice between VAR and VECM The Augmented Dickey-Fuller (ADF) test, a unit root test, is used to examine the stationarity properties of individual time series within a multivariate system. For a generic variable 𝑦𝑖,𝑡, the test is based on the following regression equation: 𝑝 𝛥𝑦𝑖,𝑡 = 𝑐𝑖 + 𝛼𝑖𝑦𝑖,𝑡−1 + (∑ 𝜃𝑖𝑗∆𝑦𝑖,𝑡−𝑗) + 𝜀𝑖,𝑡 (5) 𝑗=1 The null hypothesis 𝐻0: 𝛼𝑖 = 1 indicates the presence of a unit root, implying non- stationarity in levels. Conversely, rejection of the null suggests stationarity. The test is 13 applied sequentially to the level, first-differenced and, where necessary, second-differenced forms of each series. For variables that may exhibit exponential growth behavior (𝛼 > 1), such as price or output indices, both logged and unlogged versions were examined to assess the sensitivity of the stationarity results to functional form. This sequential procedure empirically estimates the order of integration of each variable. If every series is integrated of order one, 𝐼(1), there is reason to investigate if there is cointegration, i.e. one or more stationary linear combinations of the non-stationary series. If there is no cointegration, a Vector Autoregressive (VAR) model in first differences should be used. If there is cointegration, that suggests one or more long-run equilibrium relationship(s) and a Vector Error Correction Model (VECM) is appropriate, expressed as follows: 𝑝−1 ∆𝒀𝑡 = 𝜫𝒀𝑡−1 + (∑ 𝜞𝑖∆𝒀𝑡−𝑖) + 𝒖𝑡 (6) 𝑖=1 In equation 6, the matrix 𝜫 = 𝜶𝜷′ captures the long-run relationships, with 𝜷′𝒀𝑡−1 representing the cointegrating vectors and α the speed-of-adjustment coefficients. If 0 < 𝑟𝑎𝑛𝑘(𝜫) < 𝑘, then 𝒀𝑡 is cointegrated with 𝑟𝑎𝑛𝑘(𝜫) number of cointegrated relationships. Tests for 𝑟𝑎𝑛𝑘(𝜫) are conducted on levels data, as cointegration is inherently a property of non-differenced series. Given the results of the stationarity tests, neither a VECM nor a VAR in levels were deemed appropriate. Several variables exhibited signs of being integrated of order two, 𝐼(2), violating the assumptions required for valid cointegration analysis and thereby ruling out the use of a VECM. At the same time, a VAR model in levels requires all series to be stationary. Consequently, a VAR in first differences was selected as the most suitable specification. This choice is econometrically justified under conditions of mixed integration and further supported by model selection criteria and empirical performance. 3.4.2 Lag Order Selection The lag order (𝑝) defines how many past observations of the endogenous variables are used to explain their current values. In a VAR(𝑝) model, each variable is regressed on p of its own lags as well as p lags of all other variables in the system. The choice of 𝑝 has important 14 implications for both the statistical properties of the model and its interpretability. A lag length that is too short may result in omitted variable bias, failing to capture relevant dynamic effects. Conversely, an excessively high lag order can lead to overparameterization, reducing the efficiency of estimates and increasing the risk of multicollinearity and overfitting, particularly problematic in small samples. To determine the optimal lag length, two widely used model selection criteria were employed: the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Both criteria evaluate models based on their goodness-of-fit, as measured by the log- likelihood function, while imposing penalties for the number of parameters to avoid overfitting. They are defined as follows: 𝐴𝐼𝐶𝑝 = 2𝑘 − 2 𝑙𝑛(𝐿) (7) 𝐵𝐼𝐶𝑝 = 𝑘 × 𝑙𝑛(𝑛) − 2 𝑙𝑛(𝐿) (8) In equation (7) and (8), 𝐿 is the maximized likelihood, 𝑘 the total number of estimated parameters and 𝑛 the sample size. Lower AIC and BIC values are preferred. BIC applies a more stringent penalty due to the log-sample-size term, generally favoring more parsimonious models than AIC. In this thesis, while AIC tended to favor more complex models with additional lags, reflecting its tendency to prioritize fit over simplicity, the BIC consistently identified a single lag as sufficient. This divergence between the two is common in applied macroeconomic work. Given the relatively short time frame under investigation and the risk of reduced degrees of freedom in higher-order specifications, the more conservative BIC was adopted as the basis for final parameter selection. Accordingly, a VAR(1) model was specified. 3.5 Level Reconstruction for Interpretability Although the Vector Autoregression (VAR) model is estimated using first-differenced data to ensure stationarity, interpreting the results in differenced form can be difficult, especially in a policy context where levels are more intuitive. To enhance interpretability, forecasted values were reconstructed in levels, based on the original data transformations. 15 For variables that were log-transformed before differencing, i.e. CPI, real wages and GDP, the reconstruction utilized equation (9). Starting from the last observed level, each forecasted value was recursively re-integrated to recover its path in original units. ln(𝑌𝑡) = ln(𝑌𝑡−1) + Δ ln(𝑌𝑡) ⟹ 𝑌𝑡 = 𝑌 × 𝑒 Δ ln(𝑌𝑡) 𝑡−1 (9) For variables differenced in levels but not log-transformed, i.e. poverty and unemployment rates, equation (10) was applied. In the same manner as in the procedure above, each forecasted value was sequentially re-integrated starting from the last observed level. 𝑌𝑡 = 𝑌𝑡−1 + Δ𝑌𝑡 (10) 3.6 Isolation of Policy Effects To isolate and estimate the macroeconomic effects of the structural policy reforms introduced by President Javier Milei, a counterfactual simulation framework was employed. The fundamental goal of this approach is to compare the observed macroeconomic trajectory of Argentina in the post-reform period with a hypothetical scenario that represents how the economy might have evolved in the absence of the reforms. The model was trained using time series data from January 2017 through November 2023, deliberately excluding December 2023, the first month in which Milei's policies could have begun to influence economic variables. By restricting the training period to the pre-treatment regime, the model captures only the structural dynamics that prevailed prior to the intervention. Thus, using the model one can apply these dynamics to define the expected path of the economy in the absence of the new policy regime. The forecast of macroeconomic indicators spanned from January through December 2024. These forecasted values constitute 𝑐𝑓 the counterfactual scenario, denoted as 𝑌𝑖,?̂? , where 𝑡 ∈ {𝐽𝑎𝑛 2024, … , 𝐷𝑒𝑐 2024}. The estimated policy effect is obtained by comparing the observed values of each macroeconomic variable in 2024 (𝑌 𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑𝑖,𝑡 ), to their counterfactual counterparts: 𝑐𝑓𝑃𝑜𝑙𝑖𝑐̂𝑦 𝑒𝑓𝑓𝑒𝑐𝑡 = 𝑌 𝑜𝑏𝑠𝑒𝑟𝑣𝑒𝑑𝑖,𝑡 𝑖,𝑡 − 𝑌𝑖,?̂? (11) 16 In equation (11), the deviation between actual and forecasted values is treated as the estimated effect of the reform shock, under the assumption that no other comparably significant structural disturbances occurred during the counterfactual period. It is important to emphasize, however, that this approach does not aim to establish a causal relationship in the strict econometric sense. Rather, it provides a model-based estimate of what might have occurred in the absence of reform, conditional on structural continuity. This assumption is standard in counterfactual policy evaluation using time series models, particularly when the forecast window is limited and there is little evidence of other macroeconomic shocks during the period. 3.7 Model Validation Model validation serves as a critical step in ensuring the econometric integrity and interpretative reliability of the estimated VAR model. In this study, a deliberate focus was on three diagnostic tests, the Ljung-Box test for residual autocorrelation, the eigenvalue stability test and White’s test for heteroskedasticity. These directly correspond to the core assumptions underpinning a valid VAR framework which is serial independence, dynamic stability and homoskedasticity. 3.7.1 Structural Break Diagnostics Structural stability diagnostics assess whether short-run dynamic relationships estimated through a VAR model in first differences remain stable over time. Testing for parameter constancy allows one to evaluate whether the underlying relationships changed significantly. This ensures that the short-run model structure remains valid under potential structural breaks. To formally assess the stability of the underlying system, two standard econometric tests for structural change were conducted, the Chow test and the Cumulative Sum of Recursive Residuals (CUSUM) test. Moreover, residual plots were produced. The Chow test examines whether a structural break occurs at a known point in time, here. For a linear regression model 𝑦𝑡 = 𝛽𝑋𝑡 + 𝜀𝑡, the Chow statistic is based on comparing the residual sum of squares (RSS) before and after the break: 17 (𝑅𝑆𝑆𝑝 − (𝑅𝑆𝑆1 + 𝑅𝑆𝑆2)) 𝐹 = 𝑘 (12) (𝑅𝑆𝑆1 + 𝑅𝑆𝑆2) (𝑛1 + 𝑛2 − 2𝑘) In equation (12), 𝑅𝑆𝑆𝑝 is the RSS from the pooled regression, 𝑅𝑆𝑆1 and 𝑅𝑆𝑆2 are the RSS from subsample regressions and 𝑘 is the number of parameters. A significantly large 𝐹- statistic indicates a structural break. Different breakpoints were specified for each subsample in line with their distinct historical contexts. For the forecasting scenario, December 2023 was chosen as the breakpoint to capture the potential structural shift induced by President Javier Milei’s inauguration and the implementation of his economic reform agenda. For the counterfactual scenario, March 2020 was selected as a plausible breakpoint due to the onset of the COVID-19 pandemic, which likely introduced temporary but significant disruptions to Argentina’s macroeconomic dynamics. The CUSUM test examines the temporal stability of model parameters. If 𝜀̂ denotes the recursive residuals, the CUSUM statistic is defined as the following: 𝑡 𝜀?̂? 𝐶𝑡 = ∑ , 𝑡 = 𝑘 + 1, … , 𝑇 (13) ?̂? 𝑖=1 In equation (13), ?̂? is the standard deviation of the residuals. A plot of 𝐶𝑡 is then compared to critical bounds. 3.7.2 Search for Heteroskedasticity Evidence of significant heteroskedasticity in the residuals indicates misspecification in the variance structure of the model. This is especially relevant in the presence of macroeconomic shocks, which are likely to induce instability. In such cases, results derived from heteroskedastic VARs should be interpreted with caution or other specifications be considered. White’s test for Heteroskedasticity 18 As part of the model validation process, White’s test was employed to assess the presence of heteroskedasticity in the residuals of the estimated VAR model. Heteroskedasticity, defined as non-constant variance of the error terms that violates the classical assumption of homoskedastic disturbances in the standard linear regression framework. In a VAR context, persistent heteroskedasticity in the residuals may distort inference by biasing standard errors, reducing the efficiency of estimators and potentially masking structural instability. Mathematically, consider a generic residual series 𝜀?̂? from one of the equations in the VAR system. White’s test proceeds by regressing the squared residuals 𝜀2?̂? on the original regressors 𝑋𝑡, their squares and cross-products: 𝑘 𝑘 𝑘 𝜀2?̂? = 𝛾0 + ∑ 𝛾𝑖𝑋𝑖𝑡 + (∑ ∑ 𝛾𝑖𝑗𝑋𝑖𝑡𝑋𝑗𝑡) + 𝑢𝑡 (14) 𝑖=1 𝑖=1 𝑗=1 In equation (14), 𝑋𝑖,𝑡 are the regressors in the original model, e.g. lags of endogenous variables, 𝛾 are parameters to be estimated and 𝑢𝑡is the new error term. The null hypothesis is the following: 𝐻0: 𝐻𝑜𝑚𝑜𝑠𝑘𝑒𝑑𝑎𝑠𝑡𝑖𝑐𝑖𝑡𝑦, 𝑖. 𝑒. , 𝑉𝑎 𝑟( 𝜀𝑡 ∣ 𝑋 2 𝑡 ) = 𝜎 The alternative hypothesis is thus the following: 𝐻1: 𝐻𝑒𝑡𝑒𝑟𝑜𝑠𝑘𝑒𝑑𝑎𝑠𝑡𝑖𝑐𝑖𝑡𝑦 𝑖𝑠 𝑝𝑟𝑒𝑠𝑒𝑛𝑡 The test is based on either a Lagrange Multiplier (LM) statistic, computed as 𝐿𝑀 = 𝑛𝑅2, where 𝑅2 is the coefficient of determination from the auxiliary regression and 𝑛 is the sample size, or an F-statistic, depending on implementation. Under the null, the LM statistic is asymptotically 𝜒2-distributed with degrees of freedom equal to the number of regressors in the auxiliary regression, excluding the constant. White’s test is particularly appealing in the context of VAR models because it is general, it does not require specification of the exact form of heteroskedasticity and is robust to both 19 conditional and unconditional variance changes. In the implementation, the test is applied equation-by-equation to the residuals from each individual equation of the VAR, consistent with standard practice in multivariate time series diagnostics. Residual plots The existence of heteroskedasticity is usually, alongside statistical tests, investigated visually, especially since tests like White’s test try to prove it, rather than proving the model assumption homoskedasticity. This visual investigation is carried out through residual plots against the endogenous variables. An uneven distribution of large and small residuals indicates heteroskedasticity. 3.7.3 LM-testing As part of the model validation process, the Ljung–Box test, also referred to as a multivariate LM test in the context of VAR models, was employed to detect serial correlation in the residuals of each endogenous variable. Serial correlation violates a core assumption of the classical linear model and can undermine the consistency of parameter estimates and the validity of standard inference procedures. In a correctly specified VAR model, residuals should approximate a white noise process, that is, they should be uncorrelated over time and have constant variance. Formally, the Ljung–Box test evaluates the null hypothesis that the autocorrelations of the residuals up to a given lag 𝑝 are jointly zero, as follows: 𝐻0: 𝜌1 = 𝜌2 = ⋯ = 𝜌𝑝 = 0 This null hypothesis is put against the following alternative hypothesis: 𝐻1: ∃ 𝑗 ∈ {1, … , 𝑝} 𝑠𝑢𝑐ℎ 𝑡ℎ𝑎𝑡 𝜌𝑗 ≠ 0 The test statistic is defined as the following: 𝑝 ?̂?2𝑗 𝑄 = 𝑇(𝑇 + 2) ∑ (15) 𝑇 − 𝑗 𝑗=1 20 In equation (15), 𝑇 is the sample size and ?̂?𝑗 denotes the sample autocorrelation at lag 𝑗. Under the null hypothesis, 𝑄 asymptotically follows a 𝑥2-distribution with 𝑝 degrees of freedom. The test is applied to each residual series in the VAR model at lag 10, since 10 lags is a common diagnostic choice for monthly data. A significant 𝑝-value (< 5%) indicates autocorrelation and that the model may be mis-specified, possibly due to omitted dynamics or inappropriate lag length. Detecting autocorrelation at this stage informs whether the VAR residuals can be treated as innovation terms, a necessary condition for valid structural analysis. This step is thus crucial in ensuring that the model’s error structure conforms to the assumptions required for reliable forecasting. 3.7.4 Eigenvalue testing As part of the model validation procedure, the eigenvalue stability condition was tested to assess whether the estimated VAR model satisfies the stability requirement necessary for reliable forecasting. In the context of a VAR(𝑝) model, the system is stable or stationary in its multivariate representation, if all eigenvalues of the companion matrix lie strictly inside the unit circle in the complex plane, that is if their moduli satisfy ∣ 𝜆𝑖 ∣ < 1 for all 𝑖. The VAR(1) model this thesis uses can be recast in the following first-order companion form: 𝒁𝑡 = 𝑨𝒁𝑡−1 + 𝜼𝑡 (16) In equation (16), 𝒁𝑡 is a stacked vector of the endogenous variables, 𝑨 is the companion matrix and 𝜼 is the innovation vector. The eigenvalues of the matrix 𝑨 determine the dynamic behavior of the system. If any eigenvalue has modulus greater than or equal to one, the VAR process is said to be unstable or non-stationary. Testing the eigenvalue condition is thus a formal way to confirm that the VAR process is covariance stationary, which implies that its mean, variance and autocovariances are time- invariant. This is essential not only for theoretical coherence, but also for the empirical interpretability of the model’s outputs. 21 4. Results 4.1 Model and Parameter Selections 4.1.1 Stationarity Testing All variables exhibited non-stationarity in levels according to the Augmented Dickey-Fuller tests. This applied to both continuous economic indicators such as CPI, GDP and real wages, as well as rate-based measures like poverty and unemployment. After transformation, using log-first differences for CPI, GDP and real wages and first differences in levels for poverty and unemployment, the results remained mixed. While GDP and real wage growth appeared stationary at the 5% level, inflation, poverty and unemployment continued to exhibit signs of non-stationarity. Since not all variables became stationary after first differencing and this held true even for log-transformed series, the necessary conditions for conducting a valid Johansen cointegration test were not met. The test assumes that all variables are integrated of order one, I(1), so the presence of potential I(2) behavior among some series violates this core requirement. Applying the cointegration test under these circumstances could lead to misleading inferences about long-run relationships. Importantly, the mixed integration properties observed, where some series appear to be I(1) while others exhibit signs of I(2), also rule out the use of a standard VAR model in levels, which requires all variables to be stationary. Similarly, the VECM framework is not appropriate, as it assumes that all included variables are I(1) and cointegrated. In the presence of mixed integration orders, both approaches risk misspecification and invalid statistical inference. Instead, the analysis proceeds using a VAR model in first differences. This approach is fully justified under conditions of mixed stationarity, as long as the differenced series are approximately I(0), which is supported in part by the observed test results. The VAR in differences does not impose a long-run equilibrium structure and is robust to integration inconsistencies, making it a suitable model for short-run dynamics in this context. The first-differenced specification also benefits the economic interpretability. Log-first differences were applied to CPI, GDP and real wages to capture approximate monthly percentage changes, aligning with conventional macroeconomic interpretations of inflation 22 and growth. For poverty and unemployment, which are bounded rate variables and do not follow exponential growth, simple first differences were retained to preserve interpretability in terms of percentage point changes. Although some series remained formally non-stationary even after differencing, as shown in table 2, further differencing would compromise interpretability and remove meaningful short run variation, especially in a small sample. Given these trade-offs, the use of first differences was deemed the most appropriate strategy to ensure both empirical robustness and substantive clarity in the analysis. Table 2 – ADF tests on data from January 2017 to December 2024. Transformation Tested Variable ADF Statistic p-value Stationary? Levels CPI 4.1250 1.0000 No Levels GDP adjusted -2.4726 0.1223 No Levels Real wages -1.6563 0.4537 No Levels Poverty -1.6548 0.4546 No Levels Unemployment -1.8062 0.3774 No Log first differences Inflation -2.5578 0.1020 No Log first differences Real wages growth -3.8963 0.0021 Yes Log first differences GDP growth -3.5133 0.0077 Yes First difference in level Poverty change -2.5057 0.1141 No First difference in level Unemployment change -1.8794 0.3418 No Log first diff Poverty (log diff) -2.6385 0.0853 No Log first diff Unemployment (log diff) -1.9593 0.3047 No Second differences Inflation -4.2694 0.0005 Yes Second differences Real wages growth rate -6.3111 0.0021 Yes Second differences GDP growth rate -5.2201 0.0000 Yes Second differences Poverty change rate -8.2739 0.0000 Yes Second differences Unemployment change rate -8.2686 0.0000 Yes To confirm the reliability of the counterfactual data, ADF tests were run again using the pre- reform sample, as shown in table 3. The results largely echoed earlier findings, except CPI being non-stationary even in second difference. Real wages and GDP growth were stationary 23 in their first differences, while inflation, poverty and unemployment continued to show signs of non-stationarity. These results support the decision to use log- or first-differenced transformations across both the forecast and counterfactual models, maintaining consistency and capturing the short-term dynamics that are central to the analysis. Table 3 – ADF tests on data from January 2017 to November 2023. Transformation Variable ADF Statistic p-value Stationary? Levels CPI -1.5838 0.4917 No Levels GDP adjusted -2.2878 0.1759 No Levels Real wages -1.2194 0.6652 No Levels Poverty -1.7062 0.4279 No Levels Unemployment -1.5311 0.5181 No Log first differences Inflation 0.8514 0.9924 No Log first differences Real wages growth −3.4196 0.0103 Yes Log first differences GDP growth −3.2886 0.0154 Yes First difference in level Poverty change −2.2706 0.1816 No First difference in level Unemployment change −1.7000 0.4311 No Log first diff Poverty (log diff) -2.2914 0.1748 No Log first diff Unemployment (log diff) -1.6690 0.4472 No Second differences Inflation change rate 2.7690 1 No Second differences Real wages growth rate -6.2671 0.0000 Yes Second differences GDP growth rate -8.1297 0.0000 Yes Second differences Poverty change rate -8.8416 0.0000 Yes Second differences Unemployment change rate -7.1356 0.0000 Yes 4.1.2 Lag Order Model evaluation was conducted across four VAR specifications, varying by transformation (levels vs. log-differences) and lag structure (one or two lags). As shown in the forecasting scenario results table 4, both AIC and BIC identify the third model, a VAR(1) estimated on 24 log-differenced data, as the best-performing specification, with the lowest AIC (−30.47) and BIC (−29.66) values. Table 4 – AIC and BIC on VAR-models using data from January 2017 to December 2024. Model Transformation Lag AIC BIC 1 Levels 1 21.29 22.10 2 Levels 2 21.44 22.93 3 Log-differences 1 -30.47 -29.66 4 Log-differences 2 -30.42 -28.92 For the counterfactual scenario, shown in table 5, the same model once again outperforms all alternatives, achieving even stronger support with an AIC of −33.21 and BIC of −32.32. Although the fourth model with two lags performed similarly, it provided no additional explanatory value and increased model complexity. The chosen transformation applies log-differences to inflation, GDP and real wages, while poverty and unemployment are differenced in levels. This setup yielded the best in-sample fit and enabled consistent modeling across both analytical settings. Table 5 – AIC and BIC on VAR-models using data from January 2017 to November 2023. Model Transformation Lag AIC BIC 1 Levels 1 9.26 10.14 2 Levels 2 7.44 9.06 3 Log-differences 1 -33.21 -32.32 4 Log-differences 2 -33.04 -31.40 4.2 Policy Effects during 2024 The following results pertain to the first research question, regarding how Milei´s reforms impacted Argentina´s macroeconomic situation 2024. 25 4.2.1 Counterfactual Scenario Prices and real wages As shown in figure 1, the observed CPI rose less sharply than in the counterfactual scenario for 2024, particularly from May onwards. The counterfactual path projects a steeper, near linear increase in the price level, while the actual price level has a lower trajectory. Figure 1 – Consumer Price Index (December 2016 = 100): factual and counterfactual values for 2024. This differential is supported by the monthly policy effects reported in table 6, where inflation deviations from the counterfactual shift from positive in early 2024, e.g. +0.080 in January, to consistently negative in the second half of the year, e.g. −0.053 in November. The Root Mean Squared Error (RMSE) for inflation is 0.0467, which corresponds to a deviation of approximately 4.78% in monthly inflation, suggesting a substantial cumulative effect on the price level. In contrast, real wages followed a more volatile but ultimately positive pattern. As shown in figure 2, actual real wages initially dipped below the counterfactual in January and February. However, from April onward, observed real wages began to diverge positively, ending the year well above the counterfactual trend. 26 Figure 2 – Real Wages Index: factual and counterfactual values for 2024. Quantitatively, the RMSE for real wage growth is 0.0254, translating to a 2.58% average monthly deviation from the counterfactual. Although smaller than the inflation effect, this gap suggests meaningful consequences for purchasing power over time. Labor market and poverty As shown in figure 3, the observed unemployment rate surged during the first quarter, peaking at 7.7% in March, well above the counterfactual estimate of approximately 5.6%. However from mid-year onward, unemployment declined steadily, falling below the no- reform path from July through December. By the end of the year, the actual unemployment rate stood at 6.4%, in contrast to a near flat counterfactual trajectory of around 5.3%. Figure 3 – Unemployment rate: factual and counterfactual values for 2024. 27 Poverty dynamics, presented in figure 4, followed a similarly non-linear but more dramatic trajectory. In the first half of 2024, poverty rose sharply, reaching a peak above 52% in July – over 8 percentage points higher than the counterfactual estimate. However, the trend reversed decisively in the second half of the year. By December, the observed poverty rate had declined to approximately 38%, while the counterfactual continued rising moderately to about 47%. Figure 4 – Poverty rate: factual and counterfactual values for 2024. These patterns are numerically confirmed by the RMSE values in table 7, with poverty change registering the largest absolute policy deviation of any variable, 2.34 percentage points per month. Unemployment also shows a meaningful average monthly divergence, 0.37 percentage points, underscoring the substantial distributional effects of the policy transition. GDP Growth The trajectory of adjusted GDP in 2024 suggests a modest but meaningful deviation from the no-reform baseline. As shown in figure 5, real GDP initially declined slightly during the first half of the year, reaching a local minimum in June. However, this decline was far less pronounced than the sustained contraction projected under the counterfactual scenario. From July onward, observed GDP rebounded and grew steadily, surpassing pre-reform levels by year-end. In contrast, the counterfactual forecast continued its downward trend. 28 Figure 5 – Adjusted GDP (millions of pesos, 2004 prices): factual and counterfactual values for 2024. Quantitatively, the deviation in monthly GDP growth remained small in absolute terms, averaging approximately 0.01 log-points, but was consistently positive. The RMSE for GDP growth, at 0.0131 log-units, corresponds to an approximate percentage effect of 1.32% over the year. This modest differential, while less dramatic than the effects on prices or poverty, still reflects a measurable divergence in economic output. 4.2.2 Policy Effects Month by Month To quantify the short run macroeconomic impact of President Javier Milei’s reform package, observed outcomes for each variable in 2024 were compared against their counterfactual counterparts, as forecasted by the VAR(1) model trained on pre-reform data. The monthly difference between actual and forecasted values was estimated as the policy effect, capturing deviations possibly attributable to the reform episode. Table 6 – Estimated monthly policy effects. Month Inflation Real wages Poverty change Unemployment GDP (%) growth (%) (percentage change growth units) (percentage (%) units) 01–2024 8.03 −2.96 1.50 0.67 0.34 02–2024 2.05 0.84 1.47 0.71 0.47 03–2024 0.45 0.09 1.44 0.72 0.50 04–2024 −1.16 1.97 1.41 0.02 0.40 29 Month Inflation Real wages Poverty change Unemployment GDP (%) growth (%) (percentage change growth units) (percentage (%) units) 05–2024 −5.12 4.67 1.39 0.02 0.38 06–2024 −4.50 2.30 1.38 0.01 0.34 07–2024 −4.73 4.02 −2.97 −0.20 2.30 08–2024 −4.33 2.17 −2.97 −0.20 2.25 09–2024 −4.76 1.95 −2.98 −0.21 2.20 10–2024 −5.28 2.58 −2.98 −0.15 1.26 11–2024 −5.33 2.12 −2.98 −0.15 1.22 12–2024 −4.87 1.14 −2.98 −0.16 1.19 The estimated policy effects for 2024 reveal substantial short-run deviations from the counterfactual scenario across all five variables. Inflation initially spiked before falling below the counterfactual trajectory. Real wages fluctuated, with partial recovery mid-year. Poverty rose sharply in the first half but declined significantly thereafter. Unemployment followed a similar reversal, initially rising then steadily falling below the baseline. GDP growth was consistently higher than in the counterfactual. 4.2.3 Total Policy Effects To evaluate the overall magnitude of the policy intervention’s short-run effects, the Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) between the observed post- reform outcomes and their counterfactual trajectories for each macroeconomic variable is computed, as shown in table 7. These metrics summarize the cumulative deviation over the forecast horizon. For variables expressed in log-differences, namely inflation, real wages and GDP, RMSE is interpreted in logarithmic units, later transformed into approximate percentage deviations. In contrast, for poverty and unemployment, which are modeled in first differences in levels, RMSE is expressed in percentage points per month, representing the average absolute divergence in level changes relative to the counterfactual. 30 Table 7 – Total policy effects: mean and root mean squared error. Variable MSE RMSE Inflation 0.0022 0.0467 Real Wages Growth 0.0006 0.0254 Poverty Change 5.4539 2.3354 Unemployment Change 0.1379 0.3714 GDP Growth 0.0002 0.0131 To interpret the magnitude of these effects in more intuitive terms, the RMSE values for log- transformed variables, inflation, real wages and GDP were converted into approximate percentage deviations, using the standard exponential approximation, shown in table 8. Table 8 – Total policy effect in approximate percentages: antilog of root mean squared error. Variable % Policy Effect (antilog RMSE) Inflation 4.78 Real Wages Growth 2.58 GDP Growth 1.32 4.3 Forecast for 2025 Using the VAR(1) model trained on data from January 2017 to December 2024, a ten-month forecast was generated to project Argentina’s short-run macroeconomic trajectory from January to October 2025, shown in table 9. The forecast is made to answer the second research question, pertaining to the likely macroeconomic outlook ahead of the 2025 midterm elections. The forecast assumes no additional policy interventions or external shocks. The output is expressed in monthly log-changes, which approximate percentage growth rates. 31 Table 9 – Ten-month forecast of macroeconomic changes. Month Inflation Real wages Poverty change Unemployment GDP (%) growth (%) (percentage change growth units) (percentage units) (%) 01–2025 3.29 −0.15 −2.06 −0.16 0.39 02–2025 3.75 −0.29 −1.69 −0.13 0.34 03–2025 4.09 −0.34 −1.37 −0.11 0.29 04–2025 4.34 −0.36 −1.10 −0.09 0.24 05–2025 4.53 −0.37 −0.88 −0.07 0.19 06–2025 4.67 −0.38 −0.71 −0.06 0.15 07–2025 4.77 −0.38 −0.56 −0.06 0.12 08–2025 4.84 −0.38 −0.44 −0.05 0.09 09–2025 4.90 −0.39 −0.35 −0.04 0.07 10–2025 4.93 −0.39 −0.27 −0.04 0.05 To facilitate interpretation, the forecasted differences and log-differences were transformed into reconstructed level values, as seen in table 10. Table 10 – Ten-month forecast of macroeconomic levels. Month CPI Real wages Poverty Unemployment (%) GDP adjusted index (%) (billion pesos) 01–2025 8167 75 34 6 728 02–2025 8478 75 32 6 730 03–2025 8832 75 31 6 732 04–2025 9224 74 29 6 734 05–2025 9651 74 29 6 735 06–2025 10113 74 28 6 736 07–2025 10607 74 27 6 737 08–2025 11134 73 27 6 738 09–2025 11693 73 26 5 738 10–2025 12283 73 26 5 739 CPI 32 The forecasted trajectory of the Consumer Price Index (CPI) reflects a sustained and substantial increase in the overall price level over the ten-month forecast horizon. As shown in both table 10 and figure 6, the CPI is projected to rise from approximately 8,167 in January 2025 to 12,283 by October 2025, indicating a cumulative increase of over 50% in less than a year. This pattern suggests that while the monthly growth rate of prices remains relatively moderate, rising from a log-change of 0.033 in January to 0.049 in October, inflation – shown in figure 7 – continues to compound at a rate consistent with Argentina’s structurally persistent inflation dynamics. Figure 6 – Forecasted Consumer Price Index (December 2016 = 100) until October 2025. 33 Figure 7 – Forecasted monthly inflation rate until October 2025. Unemployment The unemployment forecast indicates a steady month-to-month decline throughout 2025, continuing the trajectory initiated in the latter part of 2024. As shown in figure 3, the unemployment rate is projected to fall from 6.08% in January to 5.42% by October, representing a total decline of approximately 0.66 percentage points over the ten-month horizon. In terms of monthly log-changes, the estimated values range from −0.157 in January to −0.040 in October, reflecting decelerating but consistently negative growth rates. 34 Figure 8 – Forecasted unemployment rate until October 2025. Real Wages Following the steep collapse in real wages during 2023–2024, the initial recovery observed in late 2024 is forecasted to lose momentum over the course of 2025. As illustrated in figure 9, the real wage index begins the year at 75.13 and gradually declines to 72.71 by October. This represents a cumulative decline of approximately 3.2%, reflecting ten consecutive months of modest negative growth. 35 Figure 9 – Forecasted Real Wages Index until October 2025. Poverty The poverty rate is projected to decline steadily and significantly throughout 2025, marking a notable reversal from the sharp spike observed in 2024. As seen in figure 10, poverty is projected to fall from 33.57% in January to 26.20% by October, amounting to a total reduction of over 7 percentage points during the forecast window. The monthly changes, ranging from −2.06 to −0.27 percentage points, suggest a front-loaded correction, with the largest improvements occurring in the first quarter of the year. Figure 10 – Forecasted poverty rate until October 2025. 36 Adjusted GDP The forecast for adjusted, real GDP is a consistent upward trajectory over the forecast horizon from January to October 2025, as shown in figure 11. Specifically, adjusted GDP rises from 728 billion pesos in January to 739 billion pesos in October, measured in 2004 prices. This represents a cumulative increase of approximately 1.5% over ten months, with monthly growth rates gradually declining from 0.39% in January to 0.05% in October. Figure 11 – Forecasted adjusted GDP (millions of pesos, 2004 prices) until October 2025. 4.4 Model Validation 4.4.1 Structural Break Diagnostics Chow breakpoint test Structural break tests based on the Chow procedure were conducted separately for the two estimation windows used in this study: one for the 2025 forecast and one for the 2024 counterfactual scenario. In each case, the Chow test was applied using a relevant breakpoint: December 2023 for the 2025 forecast, marking the inauguration of President Milei and the onset of major economic reforms, and March 2020 for the 2024 counterfactual scenario, to account for the potential structural effects of the COVID-19 pandemic. 37 For the 2025 forecast, table 11 shows statistically significant structural breaks at the 5% level for inflation, real wages growth, poverty and unemployment rate changes, with only GDP growth failing to reject the null hypothesis of parameter stability. This indicates that the macroeconomic relationships shifted notably around the time of the 2023 policy transition, supporting the treatment of this period as a structural break in the model. Table 11 – Chow tests on data from January 2017 to December 2024, assessing structural breaks in December 2023. Dependent Variable F-statistic p-value Structural break significant at 5% (Log) Inflation 6.14 0.000 Yes (Log) Real Wages Growth 2.53 0.046 Yes Poverty Change 44.80 0.000 Yes Unemployment Change 5.92 0.000 Yes (Log) GDP Growth 1.97 0.106 No For the 2024 counterfactual scenario, using March 2020 as the breakpoint, inflation, GDP Growth and poverty exhibit significant structural breaks, shown in table 12. In contrast, real wages growth and unemployment rate change appear stable over the period. Table 12 – Chow tests on data from January 2017 to November 2023, assessing structural breaks in March 2020. Dependent Variable F-statistic p-value Structural Break Significant at 5% (Log) Inflation 5.51 0.001 Yes (Log) Real Wages Growth 0.69 0.602 No Poverty Change 6.51 0.000 Yes Unemployment Change 2.09 0.090 No (Log) GDP Growth 8.02 0.000 Yes CUSUM test For the 2025 forecast, the CUSUM test rejects parameter constancy for inflation and poverty, indicating statistically significant instability in these relationships at the 5% significance level, shown in table 13. The remaining variables, real wages growth, adjusted GDP growth and unemployment rate change, exhibit no significant instability. 38 Table 13 – CUSUM tests on data from January 2017 to December 2024. Dependent variable CUSUM-statistic p-value Stable at the 5% level? (Log) Inflation 3.02 0.000 No (Log) Real Wages Growth 0.79 0.564 Yes (Log) GDP Growth 0.83 0.496 Yes Unemployment Change 0.95 0.334 Yes Poverty Change 1.80 0.003 No For the 2024 counterfactual, inflation again shows strong evidence of instability, shown in table 14. Poverty rate change also again fails the stability test at the 5% significance level. In contrast, real wages growth, adjusted GDP growth and unemployment rate change remain stable enough throughout the pre-reform period. Table 14 – CUSUM tests on data from January 2017 to November 2023. Dependent variable CUSUM-statistic p-value Stable at the 5% level? (Log) Inflation 3.03 0.000 No (Log) Real Wages Growth 0.54 0.931 Yes (Log) GDP Growth 0.70 0.714 Yes Unemployment Change 1.16 0.138 Yes Poverty Change 1.57 0.014 No 4.4.2 Search for Heteroskedasticity White’s test for Heteroskedasticity The White test results indicate clear evidence of heteroskedasticity in the forecasting scenario for all five modeled variables, shown in table 15. Each residual series, (log) inflation, (log) real wages, (log) GDP, unemployment change and poverty change, shows statistically significant test results at the 5% level, suggesting time-varying error variance across the system. 39 Table 15 – White’s test on data from January 2017 to December 2024. Variable White-statistic p-value Heteroscedasticity at 5% level (Log) Inflation 50.08 0.000 Yes (Log) Real Wages Growth 41.13 0.004 Yes (Log) GDP Growth 45.84 0.001 Yes Unemployment Change 34.02 0.026 Yes Poverty Change 31.92 0.044 Yes In the counterfactual scenario, heteroskedasticity is present in three of the five variables, shown in table 16. The residuals for (log) inflation, (log) GDP and unemployment change reject the null of homoskedasticity, while (log) real wages and poverty change do not, with p- values well above the 5% threshold. These findings suggest that volatility in residuals is more pronounced in the full sample extending into the reform period than in the pre-reform subsample. Table 16 – White’s test on data from January 2017 to November 2023. Variable White-statistic p-value Heteroscedasticity at 5% level (Log) Inflation 47.80 0.000 Yes (Log) Real Wages Growth 22.15 0.33 No (Log) GDP Growth 44.06 0.002 Yes Unemployment Change 33.29 0.031 Yes Poverty Change 19.25 0.51 No Residual plots The residual plots provide visual evidence of structural instability in several of the modeled relationships, particularly in connection with key historical events such as the onset of the COVID-19 pandemic of March 2020 and the policy shift following President Milei’s inauguration in December 2023. In the case of inflation, a pronounced spike in residuals appears directly after the reform in late 2023, as shown in figure 12, indicating that the VAR model based on pre-reform 40 dynamics fails to capture the sudden inflationary surge, suggesting a clear structural break. A similar but even stronger shift is observed in poverty rate changes in figure 13, where the residuals exhibit sharp deviations, both before and after the reform period, further indicating instability. Figure 12 – Residuals for inflation rate. Figure 13 – Residuals for poverty rate change. Real wages growth in figure 14 also display a distinct and abrupt residual drop coinciding with Milei’s inauguration, suggesting that changes in labor income dynamics were not anticipated by the model. 41 Figure 14 – Residuals for real wages growth rate. In contrast, GDP growth in figure 15 shows only a moderate disruption during the COVID-19 shock and remains comparatively stable around December 2023. Similarly, unemployment in figure 16 presents short-term fluctuations, particularly during the pandemic onset, but does not display marked residual instability in the post-reform period. Figure 15 – Residuals for GDP growth rate. 42 Figure 16 – Residuals for unemployment rate change. Taken together, the residual patterns suggest that inflation, poverty and real wages were most affected by the structural shift in late 2023, whereas GDP growth and unemployment remained more stable. 4.4.3 LM-testing The results of the Ljung-Box test (lag 10) indicate residual autocorrelation in the forecasting scenario for inflation and real wage growth, with p-values of 0.0078 and 0.0064 respectively, both significant at the 5% level, shown in table 17. This suggests that the residuals for these variables exhibit serial correlation, which may reflect dynamics not fully captured by the model. For GDP growth, unemployment change and poverty change, the test does not reject the null of no autocorrelation, indicating well-behaved residuals for these series in the forecast setting. Table 17 – Ljung-Box test on data from January 2017 to December 2024. Variable Ljung-Box Stat (Lag 10) p-value Autocorrelation at 5% Inflation 23.92 0.0078 Yes Real Wages Growth 24.49 0.0064 Yes GDP Growth 15.21 0.125 No Unemployment Change 16.90 0.0766 No Poverty Change 3.48 0.968 No 43 In the counterfactual scenario, the Ljung-Box statistics show no significant autocorrelation for any of the five variables, shown in table 18. All p-values are comfortably above the 5% threshold, with the closest being 0.059 for unemployment. These findings suggest that the residuals are generally consistent with white noise behavior in the pre-reform estimation period. Table 18 – Ljung-Box test on data from January 2017 to November 2023. Variable Ljung-Box Stat (Lag 10) p-value Autocorrelation at 5% Inflation 11.04 0.35 No Real Wages Growth 11.28 0.34 No GDP Growth 12.68 0.24 No Unemployment Change 17.78 0.06 No Poverty Change 2.34 0.99 No 4.4.4 Eigenvalue testing The eigenvalue diagnostics indicate that both the forecasting and counterfactual VAR specifications violate the stability condition, shown in table 19. In the forecasting model, all five eigenvalues have moduli greater than one, ranging from 1.20 to 4.11, confirming that the system is not dynamically stable. Table 19 – Eigenvalue modulus on data from January 2017 to December 2024. Eigenvalue Modulus Stability condition 4.11 4.11 Not stable 1.89 1.89 Not stable 1.45 1.45 Not stable 1.19 + 0.06i 1.20 Not stable 1.19 − 0.06i 1.20 Not stable Similarly, in the counterfactual model, all eigenvalues also exceed the unit threshold, with the 44 largest modulus reaching 13.71, shown in table 20. These results suggest that, in both samples, the estimated VAR models are characterized by explosive roots and do not satisfy the conditions for dynamic stability. Table 20 – Eigenvalue modulus on data from January 2017 to November 2023. Eigenvalue Modulus Stability condition -13.71 13.71 Not stable 1.65 ± 0.11i 1.65 Not stable 1.13 1.13 Not stable 1.09 1.09 Not stable 45 5. Discussion and Conclusions 5.1 Interpretation of results The empirical findings indicate that Milei’s reforms did not produce a clear short-run macroeconomic boost. This is relevant for the first research question. GDP growth showed no significant structural break and unemployment remained higher than in the no-reform scenario by end-2024, actual 6.4% vs. ~5.3% counterfactual, despite some mid-year gains. This suggests no strong evidence of immediate growth or job creation. One explanation is the limited post-reform time frame, as structural reforms typically take longer to affect macroeconomic outcomes, especially in volatile settings. This is consistent with existing research: Aguiar and Gopinath (2007) argue that large permanent shocks characterize emerging markets, making short-term gains harder to detect, while Ramey and Ramey (1995) show that high volatility can suppress growth. The initial turmoil from Milei’s reforms, including inflation spikes, poverty increases and an output dip may have weighed on early growth. Milei himself framed the reforms as short-term pain for long-term gain, which aligns with observed trends. While 2024 began with disruptions, disinflation and rising real wages emerged later in the year. Thus, the modest or insignificant first-year results do not necessarily invalidate the reforms, but rather point to delayed effects. The findings underscore that ambitious policy shifts in Latin America often take multiple quarters to yield significant results (Aguiar & Gopinath, 2007; Ramey & Ramey, 1995). Looking ahead to 2025, the uneven short-term performance of Milei’s reform agenda presents a complex challenge for the government. This regards the second research question. While the forecasted decline in inflation and recovery in real wages may bolster the administration’s economic credibility, the lingering effects of elevated unemployment and the recent memory of mid 2024’s poverty spike may continue to erode public support. The model’s forecasts suggest that although macroeconomic stabilization is underway, improvements are likely to remain modest in the near term. This dynamic has significant political implications. With midterm elections on the horizon, the administration will likely face mounting pressure to deliver not just on macro-level indicators, but also on the lived 46 economic experiences of ordinary Argentinians. Unless the trend of rising real incomes and falling poverty rates continues and is widely perceived as such, there is a tangible risk that political backlash could disrupt the long term implementation of the reform strategy. 5.2 Model Uncertainty and Robustness Several methodological limitations of the VAR analysis call for caution in interpreting the results. The post-reform sample is very short, only 13 monthly data points after Milei’s inauguration, due to the structural break in late 2023. This forced the use of a parsimonious VAR(1) in first differences, chosen via information criteria to avoid overfitting. However, this specification may omit important lag structures or long-run dynamics. With such a limited sample, even large observed changes might not reach statistical significance, reflecting low statistical power. Additionally, the underlying data generating process was unstable. Diagnostic tests, such as the CUSUM test, revealed parameter instability around the regime change, rejecting the assumption of constant coefficients. While this VAR-in-differences approach is valid under structural breaks, it cannot detect emerging long-run equilibria under the new regime. This raises concerns about sensitivity, different specifications for example more lags or level terms, might yield different impulse responses or counterfactual paths. No exogenous control variables, such as global shocks were included, so the model might over-attribute all post-2023 deviations to Milei’s reforms. Further, some key variables like unemployment, GDP and poverty were not available monthly and had to be linearly interpolated from quarterly or semi-annual data. This smoothing likely understates true volatility and may mute sharp changes, such as the mid-2024 poverty spike and recovery. As a result, impulse responses and forecast errors might appear smaller than they were, potentially underestimating short-run reform impacts. In sum, the analysis is limited by short data horizons, structural instability, absence of exogenous controls and interpolation bias. These factors imply that the findings are suggestive, not definitive. Small adjustments to model design or added data could shift conclusions. Robustness checks, such as using alternative lag lengths or including a December 2023 break dummy, are therefore essential. While this analysis reflects the best 47 effort given the constraints, substantial model uncertainty remains regarding the size and timing of the reform effects. 5.3 Political and Economic Contextualization Placing the quantitative findings in Argentina’s broader context clarifies why the measured effects of Milei’s reforms were limited or slow to emerge. Politically, Milei operated under constraints, his party held only a minority in Congress, leading to delays and dilution of key reforms such as the omnibus bill, which was not passed until mid 2024. As a result, the actual reform program implemented was less extensive and slower than initially planned, likely muting the expected macroeconomic effects. Partial implementation, especially in areas like labor deregulation, means large short-run improvements in employment or investment were unlikely. Milei’s own policy mix also produced mixed early outcomes. Aggressive austerity and sudden deregulation contributed to a Q1–Q2 2024 spike in unemployment and poverty as public sector jobs were cut and subsidies removed. Yet his strong anti-inflation stance and signals of eventual dollarization may have later anchored expectations, contributing to disinflation, from over 20% monthly inflation in late 2023 to low single digits by the end of 2024. International influences also played a major role. The IMF endorsed Milei’s austerity with a $20 billion support package in late 2024, accelerating disbursements to bolster reserves. This external backing likely stabilized the currency and boosted investor confidence, effects not separately captured by the VAR model. Simultaneously, exogenous shocks blurred the reforms’ true impact. A severe 2023 drought reduced GDP by an estimated 3% and its reversal in 2024 may explain part of the modest rebound. Similarly, global interest rates and commodity price shifts, soy exports and oil prices for example, affected trade balances and inflation independently of Milei’s policies. While the VAR model controlled for preexisting trends, it could not isolate all external factors. Thus, the moderate reform effects observed in 2024 reflect a mix of domestic policy, institutional constraints and global conditions. For example, the late 2024 improvements in real wages and poverty may stem as much from disinflation helped by a stable exchange rate 48 and commodity trends as from structural reforms. As such, any assessment of Milei’s program must factor in these confounding influences. Argentina’s 2024 macroeconomic trajectory cannot be explained by policy changes alone. Instead, it reflects a complex interplay between Milei’s reform agenda, institutional and political limits, IMF involvement and external shocks, an interplay the VAR model seeks to approximate, although with noted limitations. 5.4 Policy Implications The findings offer several important implications for policymakers in Argentina and other high-volatility emerging economies. One key lesson is the importance of managing the transition period during major reforms. While Milei’s shock therapy delivered significant achievements by the end of 2024, most notably sharp disinflation and a restored primary fiscal balance, it also entailed considerable short-run costs. Poverty exceeded 50% in mid- 2024, around 8 percentage points higher than the counterfactual, and unemployment surged early on. These effects later eased as inflation declined and confidence improved, but the data highlight that stabilization policies can intensify social hardship in the short term. This underscores the need for short run safety nets. Policymakers should accompany aggressive reforms with targeted support like temporary cash transfers, unemployment insurance or emergency subsidies to mitigate early welfare losses. In Argentina’s case, the lack of automatic stabilizers meant the poorest were hit hardest, which future reformers could avoid for both humanitarian and political reasons. Maintaining public support through the turbulence is critical, cushioning measures can help sustain reform momentum until long term benefits emerge. Another implication concerns the timing and sequencing of reforms. Milei implemented multiple reforms simultaneously, fiscal tightening, monetary contraction, deregulation, which maximized the immediate shock. While this “shock therapy” approach yielded rapid results, such as ending inflation and eliminating the fiscal deficit, it also caused sharp disruptions to incomes and employment. Policymakers must weigh this trade-off. In crisis conditions, swift and comprehensive reforms may be essential. Otherwise, a more gradual, sequenced program with pre-arranged compensations like aligning subsidy removals with transfers, could ensure smoother adjustment. 49 Credibility and external support also proved vital. Milei’s fiscal orthodoxy, reinforced by IMF involvement, helped anchor expectations and stabilize inflation. The IMF’s $20 billion support package provided needed reserve buffers and financial credibility. This suggests that pairing domestic reforms with external anchors, such as IMF oversight, sovereign bond commitments or rule-based policy frameworks, can strengthen reform effectiveness by reducing uncertainty and building trust. Lastly, rigorous policy evaluation is essential. This study’s use of a counterfactual VAR approach illustrates how headline outcomes, like inflation or growth must be assessed against what would have happened otherwise. Policymakers should institutionalize such evaluative thinking. This includes improving data collection, especially high-frequency social indicators, ensuring transparent progress reporting and establishing independent evaluation mechanisms. These tools can guide timely course corrections and reinforce both the political and economic sustainability of reforms. The Argentine case reinforces classic policy lessons: align reform timing with crisis needs, prioritize social safety nets, secure credibility through external anchors and ensure continuous, data-driven evaluation. 5.5 Suggestions for Future Research Given this study’s limitations, several directions for future research could deepen understanding of Milei’s reform impact. First, re-estimating the model with a longer post- reform sample is crucial. As more data become available from 2025 onward, researchers can assess whether early positive signs, disinflation, real wage gains, poverty reduction, evolve into lasting improvements. A longer horizon will also clarify second year effects, such as potential rebounds in growth and investment. Revisiting the VAR, or a similar model, over time could determine whether Milei’s promised “long-term gains” materialize and if early sacrifices were justified. Second, applying alternative empirical methods would enhance robustness. While this thesis used a counterfactual VAR, other techniques could address its limitations. Structural VARs (SVARs), for example, could identify reform shocks more directly, using exogenous instruments like Milei’s election or the December 2023 reform bill, and impose theoretical 50 restrictions to isolate effects. Another option is local projections, which estimate impulse responses without relying on a fixed VAR structure and are more resilient to model misspecification. In addition, micro level approaches such as difference-in-differences (DiD) could uncover distributional and sectoral effects. By exploiting heterogeneity, for example across industries affected differently by deregulation or subsidy cuts, researchers could analyze employment, investment or consumption at the household or firm level. This would reveal who gained or lost from the reforms, such as whether benefits were concentrated among exporters or high- skilled workers. Future studies should also broaden the set of macroeconomic outcomes analyzed. This thesis focused on inflation, GDP, unemployment, poverty and real wages, but other indicators like investment, exchange rate stability or industrial productivity, could yield insights into reform mechanisms. Investment trends would show whether Milei’s agenda boosted business confidence. Exchange rate analysis, especially using high-frequency data, could assess credibility effects, such as the narrowing of the parallel market gap. Event studies might track market reactions to key announcements. Sector-specific outcomes like manufacturing output or productivity would help gauge supply side responses. Social metrics beyond poverty, such as inequality or consumption patterns could illuminate whether the reforms had inclusive effects. These inquiries would benefit from more data over time and also cross sectional variation to improve identification. Future research can better evaluate Milei’s reforms by extending the timeline, applying varied methodologies and examining a broader array of indicators. This expanded approach will help determine the true impact of Argentina’s radical economic shift and offer policy lessons for other volatile emerging markets. 51 References Aguiar, M. & Gopinath, G. (2007). “Emerging Market Business Cycles: The Cycle Is the Trend.” Journal of Political Economy, 115(1), pp. 69–102. DOI: 10.1086/511283. Ames, B., 1987. Political Survival: Politicians and Public Policy in Latin America. Berkeley: University of California Press. Associated Press (2024). In key milestones for President Milei, Argentina secures IMF deal and ends most capital controls. AP News. Available at: https://apnews.com/article/argentina- economy-inflation-milei-imf-consumer-prices-currency- 8ca2cc2b1a2b025e8913f8ae85cd7170 [Accessed 2025-05-16] Brender, A. & Drazen, A. (2005). “Political Budget Cycles in New versus Established Democracies.” Journal of Monetary Economics, 52(7), pp. 1271–1295. Cermeño, R., Grier, R. & Grier, K. (2010). “Elections, Exchange Rates, and Reform in Latin America.” Journal of Development Economics, 92(2), pp. 166–174. Drazen, A., (2000). The Political Business Cycle after 25 Years. NBER Working Paper No. w7352, National Bureau of Economic Research. Available at: https://doi.org/10.3386/w7352 [Accessed 2025-05-15] Duncan, R. and Martínez-García, E., (2018). New perspectives on forecasting inflation in emerging market economies: An empirical assessment. Globalization and Monetary Policy Institute Working Paper No. 338, Federal Reserve Bank of Dallas. Available at: https://doi.org/10.24149/gwp338 [Accessed 2025-05-16]. Falcone, M. (2024). One Year of Milei: Stabilization, a Balanced Budget and Deregulation in Argentina. Econlib (Economic Liberty Blog). Available at: https://www.econlib.org/one- year-of-milei-stabilization-a-balanced-budget-and-deregulation-in-argentina/ [Accessed 2025-05-16] Frankel, J.A., (2017). How to cope with volatile commodity export prices: Four proposals. CID Faculty Working Paper No. 335, Center for International Development at Harvard University. 52 Green, R. (2024). Milei’s radical reforms risk rolling back labour rights and rule of law in Argentina. International Bar Association News. Available at: https://www.ibanet.org/mileis- radical-reforms-risk-rolling-back-labour-rights-and-rule-of-law-in-argentina [Accessed 2025- 05-18] Harguindeguy, S. (2024). From Milei’s zero fiscal deficit towards a stabilisation plan to eradicate inflation: why now? Real Instituto Elcano Analysis Paper. Available at: https://www.realinstitutoelcano.org/en/analyses/from-mileis-zero-fiscal-deficit-towards-a- stabilisation-plan-to-eradicate-inflation-why-now/ [Accessed 2025-05-18] Ho, G. and Mauro, P., (2016). Growth: the past, the present, and the future. In: S. M. Ali Abbas, A. Pienkowski and K. Rogoff, eds. Macroeconomics of Austerity in Latin America: Recent Lessons. Washington, DC: International Monetary Fund, pp. 19–38. Mandalinci, Z. (2017). “Forecasting Inflation in Emerging Markets: An Evaluation of Alternative Models.” International Journal of Forecasting, 33(4), pp. 1082–1104. Mihaljek, D. & Saxena, S. (2010). “Wages, Productivity and ‘Structural’ Inflation in Emerging Market Economies.” BIS Papers No. 49, Bank for International Settlements. Nordhaus, W.D. (1975). “The Political Business Cycle.” Review of Economic Studies, 42(2), pp. 169–190. Paqué, K.-H. & Holtzmann, H.-D. (2024). One year of Javier Milei’s economic policy: Successes, challenges and recommendations for action. Friedrich Naumann Foundation. Available at: https://www.freiheit.org/one-year-javier-mileis-economic-policy [Accessed 2025-05-19] Ramey, G. & Ramey, V. (1995). Cross-Country Evidence on the Link Between Volatility and Growth. American Economic Review, 85(5), pp. 1138–1151. Shi, M. & Svensson, J. (2006). “Political Budget Cycles: Do They Differ Across Countries and Why?” Journal of Public Economics, 90(8–9), pp. 1367–1389. Sims, C.A. (1980). “Macroeconomics and Reality.” Econometrica, 48(1), pp. 1–48. 53 Soltys, M. (2024). Five key points from Milei’s “omnibus” reform package. Buenos Aires Times. Available at: https://www.batimes.com.ar/news/argentina/five-key-points-of-mileis- reform-package.phtml [Accessed 2025-05-18] Werner, A. (2025). Milei in 2025: Between Argentina’s mid-term elections and the IMF. Peterson Institute for International Economics (PIIE) RealTime Economics Blog. Available at: https://en.mercopress.com/2025/01/28/milei-s-2025-challenges-between-argentina-s-mid- term-elections-and-the-imf [Accessed 2025-05-16] World Bank, (2020). Considering Labor Informality in Forecasting Poverty and Inequality. Washington, DC: World Bank. Available at: https://documents1.worldbank.org/curated/en/099516206222334188/pdf/IDU0847415ac0dc7 8040150bdd90714b239a7bb9.pdf [Accessed 2025-05-20]. 54