Iranian Journal of Accounting, Auditing and Finance

Iranian Journal of Accounting, Auditing and Finance

Do Machine-Learning Models Beat Simple Benchmarks in Forecasting Next-Day Absolute Price Changes? Evidence from the Tehran Stock Exchange

Document Type : Original Article

Authors
1 PhD candidate in Accounting, Alborz Campus, University of Tehran
2 Department of Accounting, Ma.C., Islamic Azad University, Mashhad, Iran
3 Master of financial management, Science and Research Branch, Islamic Azad University, Tehran, Iran.
10.22067/ijaaf.2026.98996.1669
Abstract
Forecasting the size of daily price movements matters for risk management, yet evidence on how well machine-learning models perform out of sample in emerging markets is scarce. This study asks whether machine-learning models can forecast the next-day absolute percentage change of the Tehran Stock Exchange (TSE) index, defined as the absolute open-to-close percentage change and used as a magnitude-based proxy for realized volatility. Using 1,512 daily observations and a strictly chronological 80/20 split, we compare five approaches (K-Means clustering, a rule-based classifier, a decision tree, gradient boosting regression (GBR), and support vector regression (SVR)) with five benchmarks, including a HAR-type regression and GARCH(1,1). Accuracy is measured by the out-of-sample R² relative to the historical mean (R²OS). In the final test block, which is calmer than the training period, R²OS is 10.7% for HAR, 7.8% for GARCH, 6.4% for GBR, and −0.7% for SVR; in a rolling-origin evaluation with 756 pooled forecasts it is 13.8%, 12.6%, 11.2%, and 6.8%. No machine-learning model using the eight lagged predictors improves on HAR, which additionally uses a 22-day average and is significantly better than GBR (p = 0.029) and SVR (p = 0.006) in the rolling evaluation. The decision tree is less accurate than always predicting the low-movement class (69.0% against 75.6%), and K-Means clusters carry no information on next-day movement. Same-day information, in contrast, explains over 90% of the variance mechanically, which cautions against high R² values obtained with contemporaneous inputs.
Keywords
Subjects

Send comment about this article
Enter Name.
Enter a valid email address.
Enter a vaid affiliation.
Enter comments (At leaset 10 words)
CAPTCHA Image
Enter Security Code Correctly.

Articles in Press, Accepted Manuscript
Available Online from 18 November 2026