HSE Researchers Demonstrate Effectiveness of Machine Learning in Forecasting Inflation

Perm, Russia

© iStock

Inflation is a key indicator of economic stability, and being able to accurately forecast its levels across regions is crucial for governments, businesses, and households. Tatiana Bukina and Dmitry Kashin at HSE Campus in Perm have found that machine learning techniques outperform traditional econometric models in long-term inflation forecasting. The results of the study focused on several regions in the Privolzhskiy Federal District have been published in HSE Economic Journal.

Inflation forecasting is crucial for the economy, particularly since Russia's transition to an inflation targeting regime in 2014. This means that the Bank of Russia sets specific inflation targets and employs a range of tools to meet them.

Various data is used to predict inflation, including the consumer price index, the unemployment rate, exchange rates, and the Central Bank rate. To systematise this information for forecasting, economists at the HSE Campus in Perm used data from the Unified Interdepartmental Statistical Information System.

The researchers' main objective was to determine which model predicts regional inflation more accurately: traditional econometric time series models or more recent machine learning methods. The study analysed data from 14 regions in the Privolzhskiy Federal District from January 2010 to December 2022. R Studio and Python were used for the analysis: time series forecasting was performed in R Studio, while machine learning models, including support vector machines, gradient boosting, and random forests, were implemented in Python. The forecasts were conducted on test samples, which helped prevent model overfitting and provided more accurate estimates.

The authors employed a cross-validation method using test samples of equal size. This approach allows models to be trained on data from one period and tested on data from another, ensuring stability and accuracy of the forecasts.

Tatiana Bukina

'To ensure accurate performance of machine learning methods, it is essential to select the optimal hyperparameters for the models. Hyperparameters differ from other model parameters in that they are set before training begins and define the model's specifications. Cross-validation is employed to select the optimal hyperparameters. When cross-validating time series, the training data precedes the test data without overlap, unlike in standard data validation,' according to Tatiana Bukina, Associate Professor, Faculty of Computer Science, Economics, and Social Sciences, HSE Campus in Perm.

The study found the gradient boosting model to be the most accurate of all machine learning models considered for predicting regional inflation. It delivers more accurate forecasts than autoregressive models over more time horizons. Thus, at forecasting horizons of 3, 6, 21, and 24 months, the gradient boosting model outperforms the basic AR(1) model by 20.3%, 16.2%, 72.5%, and 77.7%, respectively. The AR(1) model, a statistical tool for analysing and predicting time series, assumes that the current value of a series depends on its previous value plus a random error.

The random forest model and the support vector machine also demonstrated accurate forecasts over the long horizons of 21 and 24 months, outperforming the AR(1) model by 72.5% and 77.7%, respectively. A random forest combines multiple decision trees to enhance the accuracy and stability of forecasts, and then uses regression to average the predictions or select the most frequent value. The support vector machine identifies the optimal line that separates the data while minimising classification errors.

Inflation and its forecast for the Perm Region

© Bukina T., Kashin D. (2024) Prognozirovanie regional'noy inflyatsii: ekonometricheskie modeli ili metody mashinnogo obucheniya? HSE Economic Journal, vol. 28, no 1, pp. 81-107.

According to the authors, their results confirm that machine learning methods can be effective for forecasting inflation across various time horizons. 

Tatiana Bukina notes, 'Our research has demonstrated that machine learning provides more reliable tools for long-term forecasts. However, traditional econometric models continue to play a crucial role in short-term forecasts and should not be entirely excluded from analysts' toolkits. Combining econometric modelling with machine learning methods can significantly improve the accuracy of regional inflation forecasts. This is particularly important in an environment characterised by high uncertainty and rapid changes in economic conditions.'

The study also highlighted the specific characteristics of inflation forecasting for different regions. For example, in machine learning models, inflation seasonality was observed only in the Perm, Nizhny Novgorod, Penza, and Saratov regions. In the Republic of Tatarstan, the specific month for which the forecast was calculated proved to be a significant factor.

In the random forest model, the average inflation value for the previous three months emerged as a significant factor for the Republic of Mordovia, Nizhny Novgorod and Ulyanovsk regions, and the Chuvash Republic.

Each region has unique characteristics related to its economic structure, natural resources, and geographical location. These factors account for the variations in inflation dynamics and key macroeconomic indicators.