📈 economics

Evaluating Machine Learning and Econometric Models for Inflation Forecasting: Evidence from a Sub-Saharan African Panel

This study evaluates various machine learning and econometric models for inflation forecasting across eight Sub-Saharan African economies and finds that, due to structural breaks and high volatility, no complex model statistically outperforms a simple naive persistence baseline, suggesting that regional policymakers should maintain simple models as active benchmarks.

Original authors: Baffour Osei, Abigail Boatemaa, Isaac Nyame, Gabriel Osei Forkuo

Published 2026-07-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Baffour Osei, Abigail Boatemaa, Isaac Nyame, Gabriel Osei Forkuo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Great Prediction Race: Why the Simplest Guess Often Wins

Imagine you are trying to guess the weather for next Tuesday. You could pull out a supercomputer, analyze satellite images, measure wind speeds, and run complex climate models. Or, you could just look out the window and say, "It's going to be about the same as it is today." In the world of economics, this is the ultimate showdown between high-tech complexity and simple common sense. Economists have spent decades building massive, intricate models to predict inflation—the rate at which prices for things like bread, fuel, and rent go up. They use fancy math and machine learning, hoping to find hidden patterns that will help governments set interest rates and keep families from losing their savings. But there is a stubborn, old-school rival in this race: the "random walk." This is the idea that the best guess for tomorrow's price is simply today's price. It sounds silly, like guessing a coin flip, but history shows that this simple trick is incredibly hard to beat, especially in places where the economy is wild and unpredictable.

This paper dives into that exact battle, but with a twist: it tests these ideas in Sub-Saharan Africa, a region where inflation can swing wildly like a pendulum. The researchers gathered data from eight different countries over more than 50 years, creating a massive digital playground to see if modern "smart" computers (Machine Learning and Deep Learning) could finally outsmart the simple "look-at-yesterday" method. They didn't just look at who had the lowest error number; they used strict statistical tests to see if any fancy model was actually better, or if they were just getting lucky.

The Race: High-Tech vs. The "Yesterday" Strategy

The authors set up a grand experiment involving eight Sub-Saharan African economies: Ghana, Nigeria, Kenya, Côte d'Ivoire, South Africa, Senegal, Tanzania, and Uganda. They built a dataset of 424 snapshots of economic life, stretching from 1972 to 2024. Think of this as a giant training camp where they taught eight different "coaches" how to predict inflation.

The coaches were a mix of old-school and new-school:

The Naive Coach (Random Walk): This coach has no brain. It just says, "Next year's inflation will be exactly what it was this year."
The Econometric Coaches: These are the traditional experts who use linear equations and fixed rules based on economic theory.
The Machine Learning Coaches: These are the high-tech wizards. They include Random Forest (a team of decision trees), XGBoost (a super-optimized tree booster), and Elastic Net (a smart filter that picks the best clues).
The Deep Learning Coach (LSTM): This is the most complex one, a type of artificial intelligence designed to remember long sequences of events, like a human remembering a story.

The researchers split their data into three parts: a training set (to teach the coaches), a validation set (to tune them), and a test set (the final exam from 2018 to 2024). They wanted to see if the fancy AI could beat the simple "look-at-yesterday" coach.

The Shocking Result: The Simple Coach Holds the Line

When the final exam results came in, the outcome was a massive shock to the idea that "bigger and smarter is always better."

The Naive Coach (Random Walk) was statistically undefeated. It achieved the lowest error rate of any model, with a Test RMSE (a measure of how far off the guess was) of 4.435. More importantly, when the researchers ran strict statistical tests (called Diebold-Mariano tests) to see if any other model was truly better, the answer was a resounding no. None of the complex models could prove they were statistically superior to the simple "look-at-yesterday" guess.

Here is how the other coaches fared:

The Machine Learning Wizards (Random Forest and XGBoost): These were the best of the complex bunch. Random Forest came in with a Test RMSE of 4.849, and XGBoost had 5.825. While they were close, the statistical tests showed they were essentially tied with the Naive Coach. They didn't lose, but they didn't win either; they achieved statistical parity. Random Forest was the only model that could claim a statistical tie with the simple baseline, but it did not surpass it.
The Traditional Experts (Pooled OLS and Panel Fixed Effects): These old-school linear models did poorly. They were significantly beaten by the Naive Coach, with errors of 6.781 and 6.612 respectively. The paper suggests that trying to force a straight-line relationship onto a volatile, messy economy just doesn't work.
The Deep Learning Star (LSTM): This was the most dramatic story. When the AI was trained for too long (350 epochs) without a safety brake, it suffered from "overfitting." It memorized the training data so perfectly that it failed the test, scoring a terrible 7.741. However, when the researchers added "early stopping" (a technique to stop training before the AI gets too confused), its score improved to 6.162. This was a huge improvement, but it still didn't beat the simple Naive Coach. It just became "statistically indistinguishable" from it—meaning it was good enough to tie, but not good enough to win.

Why Did the Smart Models Fail?

The authors offer a few reasons why the high-tech models couldn't crack the code in this specific region.

First, the data is too small and too noisy. The training set had only 304 observations. For a super-complex AI like an LSTM, that's like trying to learn a language by reading only a few pages of a dictionary. The models get confused and start memorizing the noise rather than learning the rules.

Second, the economy keeps changing the rules. The test period (2018–2024) included massive shocks like the COVID-19 pandemic and global commodity price spikes. The "simple" model just looks at the most recent number, which adapts instantly to these shocks. The complex models, however, try to find a pattern based on the past (pre-2009 data), which no longer applies when the world changes suddenly. It's like trying to navigate a storm using a map from a sunny day; the map is too detailed, but it's the wrong map.

Third, the "clues" might not be that helpful. The models were fed data like interest rates, government spending, and oil prices. But in these economies, inflation seems to be driven mostly by its own momentum and exchange rates, making the other clues less useful. The simple model ignores all those extra clues and just rides the momentum, which turns out to be the winning strategy.

What This Means for the Future

The paper concludes that for policymakers in Sub-Saharan Africa, the message is clear: Don't throw away the simple tools just because you have fancy new ones.

While the complex models (like Random Forest and the corrected LSTM) are impressive and can match the simple model's performance, they haven't proven they can beat it. In fact, the simple "persistence" model is so robust that it should remain the "gold standard" or the "baseline" that all new models must beat before being trusted.

The authors suggest that if central banks want to use these fancy AI tools, they should do so cautiously. They recommend keeping the simple model running in the background as a reality check. If a complex model can't statistically prove it's better than just guessing "tomorrow will be like today," then it's probably not worth the extra cost and complexity.

In the end, this study adds another chapter to a long-running mystery in economics: in a world full of chaos and sudden changes, sometimes the smartest thing you can do is just look at what happened yesterday and assume it will happen again. The "simple" isn't always the best, but in the volatile markets of Sub-Saharan Africa, it's currently the champion.

Technical Summary: Evaluating Machine Learning and Econometric Models for Inflation Forecasting in Sub-Saharan Africa

Problem Statement
Policymakers in Sub-Saharan Africa (SSA) face increasing pressure to adopt machine learning (ML) and deep learning (DL) methods for inflation forecasting. However, the regional literature rarely benchmarks these complex approaches against simple persistence models (random walks). Furthermore, existing studies often rely on single-country time series with limited sample sizes (approx. 50 annual observations), which are insufficient for training high-capacity models without severe overfitting. There is a critical gap in understanding whether increased model complexity yields statistically defensible gains in data-constrained, high-volatility SSA environments when rigorously compared to a naive baseline.

Methodology
The study constructs a reproducible, annual panel dataset covering eight SSA economies (Ghana, Nigeria, Kenya, Côte d'Ivoire, South Africa, Senegal, Tanzania, and Uganda) from 1972 to 2024 ( $N=424$ ). Data were extracted programmatically via APIs from the World Bank (WDI) and FRED to ensure auditability.

Data Preprocessing: The pipeline strictly prevents cross-sectional data leakage. Imputation (forward/backward fill) and lag generation (1st and 2nd order) were performed within country groups before the chronological split. Features were standardized using only training set statistics.
Model Suite: Eight forecasting approaches were evaluated:
1. Naive Benchmark: A country-specific random-walk persistence model (Atkeson-Ohanian).
2. Linear Econometrics: Pooled OLS and Panel Fixed Effects (within estimator).
3. Machine Learning: Elastic Net (L1/L2 regularized), Random Forest, and XGBoost.
4. Deep Learning: A Panel LSTM (Recurrent Neural Network) evaluated with and without early stopping.
Evaluation Framework: Models were trained on a strict chronological split (Train: $\le$ 2009; Validation: 2010–2017; Test: 2018–2024). Predictive accuracy was assessed using RMSE and MAE, but statistical significance was determined via pairwise Diebold-Mariano (DM) tests on squared-error loss differentials.
Robustness Checks: The study employed HAC-Newey-West/Harvey-Leybourne-Newbold (HLN) corrected variance estimators, Holm-Bonferroni adjustments for multiple comparisons, strictly lagged-only predictor specifications (removing contemporaneous exogenous variables), expanding-window recursive re-estimation, and LSTM sequence-length sensitivity analysis.

Key Results
The empirical analysis yields a definitive "random-walk puzzle" result for the SSA context:

Dominance of the Naive Benchmark: No estimated model (econometric, ML, or DL) achieved a statistically significant improvement in out-of-sample forecast accuracy over the naive persistence benchmark. The naive model achieved the lowest point-estimate Test RMSE (4.435) and MAE (2.510).
Performance of Tree Ensembles: Random Forest and XGBoost achieved statistical parity with the random walk (DM p-values of 0.538 and 0.130, respectively) but did not surpass it. Notably, Random Forest significantly outperformed XGBoost on the full panel, a reversal of trends seen in smaller interim analyses.
Failure of Linear and Uncorrected DL Models: Traditional linear models (Pooled OLS and Panel Fixed Effects) and the uncorrected LSTM (trained for 350 epochs without early stopping) were significantly dominated by the naive benchmark at the 5% level.
Impact of Early Stopping: Implementing early stopping on the LSTM reduced Test RMSE by 20.4% (from 7.741 to 6.162), moving it from significantly worse than the benchmark to statistically indistinguishable from it. However, it did not yield a competitive advantage over the tree-based ensembles.
Robustness: These findings held under stricter HAC-HLN variance estimation and Holm-Bonferroni corrections (where only the difference between the naive benchmark and Panel Fixed Effects remained significant). Results were also consistent when using a strictly lagged-only feature set, confirming the findings are not artifacts of using contemporaneous information unavailable at the forecast origin.

Key Contributions

Empirical Benchmarking: This is among the first studies to systematically include the naive persistence benchmark in SSA-specific inflation forecasting, challenging the assumption that complex models inherently outperform simple baselines in this region.
Reproducible Pipeline: The study provides a fully auditable, API-driven data extraction and preprocessing pipeline that avoids cross-sectional leakage, setting a transparency standard for applied macroeconometric work in data-scarce environments.
Model Class Expansion: It extends the "random-walk puzzle" literature beyond Phillips-curve and VAR models to include modern tree-based ensembles and recurrent neural networks, demonstrating that the puzzle persists even with these advanced architectures in annual SSA data.
Methodological Rigor: The application of formal Diebold-Mariano testing with multiple-comparison corrections and robustness checks (recursive re-estimation, lagged-only specs) provides a more rigorous evaluation framework than point-estimate comparisons alone.

Significance and Claims
The paper claims that its findings align with the well-documented "random-walk puzzle" in global inflation forecasting, suggesting that structural breaks, high volatility, and modest sample sizes pose substantial challenges for highly parameterized forecasting architectures in SSA.

The authors conclude that regional central banks and finance ministries should maintain simple persistence models as active baselines when evaluating investments in complex forecasting infrastructure. The study advises against over-investing in complex models without first demonstrating they clear the low bar set by persistence. It suggests that while tree-based ensembles can achieve parity with simple models, they do not offer a statistically significant superiority in this specific data regime. Consequently, institutions should prioritize formal pairwise predictive accuracy tests (like Diebold-Mariano) over point-estimate metrics when validating new forecasting tools.