Optimized machine learning for BIST profitability prediction

Abstract

Forecasting firm profitability from financial-panel data requires flexibility in prediction and time-respecting validation. In this study, we forecast Return on Assets (ROA), Return on Equity (ROE) and Net Profit Margin (NPM) for technology firms listed on Borsa Istanbul exactly one-quarter ahead using 1,200 firm-quarter pairs from 29 firms between 2007 and 2024. We first evaluate eight untuned baseline models with a chronological holdout and purged expanding-window validation with fold-local preprocessing. Gradient Boosting, XGBoost and LightGBM are tuned with Grid Search, Random Search, Optuna/TPE, original Survivor Optimizer (SO), and proposed Surrogate-Guided Stagnation-Adaptive Survivor Optimizer (SGSA-SO), with 30 expensive evaluations and 10 independent runs for each. SGSA-SO has the lowest mean inner-validation MAE for seven of nine model-target tasks and ranks in the top-two for all nine. Holm-adjusted tests do not indicate that it is universally superior to Optuna. Relative to the untuned version of the same selected model family, HPO lowers held-out MAE and MAPE for all three targets; RMSE also improves for ROA but is essentially unchanged for ROE and slightly higher for NPM because of extreme observations. Training-only selection yields XGBoost–SGSA-SO for ROA, GB–SGSA-SO for ROE, and GB–Grid for NPM, with held-out MAE/RMSE/MAPE of 8.264/12.384/312.3%, 32.937/295.949/258.5%, and 249.010/2813.228/2164.4%, respectively. We report MAPE with coverage, but do not use it to select the model. The results support budget-matched HPO, but highlight the need for strict temporal validation.