Running summarizing tools on a new article

Check the paper below, then start its synthesis.


Statistical Accuracy Versus Trading Utility: A Multi-Horizon Evaluation of Generative-Ai-Augmented Deep Learning on the Kse-100 Index

Zaka Ullah Saif (University of Central Punjab), Saira Gillani (University of Central Punjab), Anila Gulzar Toor

International Journal of Innovations in Science and Technology
DOI: 10.33411/ijist/1918 (Decision Sciences)
License: CC BY

Abstract: Forecasting financial markets is increasingly approached through deep learning, with recent work proposing generative augmentation as a means to improve predictive accuracy under the high noise and non-stationarity that characterize emerging-market equities. This study evaluates whether such accuracy gains translate into trading utility on the Karachi Stock Exchange 100 (KSE-100) index, using daily data from February 2008 to February 2024 (4,176 observations) spanning three crisis episodes. Eleven modeling configurations: Ridge regression, Random Forest, XGBoost, LSTM, GRU, BiLSTM, PatchTST, and a β-variational-autoencoder augmentation of each deep-learning backbone are compared under an identical expanding-window walk-forward protocol with thirteen non-overlapping test folds, four forecast horizons (1, 5, 10, and 20 days) and a cost-aware backtest applying 5 bps fees plus 2 bps slippage. Performance is reported on both statistical (MAE, RMSE, Diebold–Mariano) and economic (Sharpe ratio, CAGR, maximum drawdown, turnover) criteria. Prior evaluations of GenAI augmentation in finance assess a single backbone at a single horizon without cost-aware trading validation. No such multi-horizon, multi-backbone evaluation has previously been reported for the KSE-100. PatchTST+GenAI achieves the lowest mean MAE (0.0450) and the lowest mean RMSE (0.0566) across horizons, with Diebold–Mariano tests confirming statistically significant accuracy gains over its unaugmented counterpart at horizons of 5, 10 and 20 days. However, this accuracy advantage does not translate to trading utility: PatchTST+GenAI records a Sharpe ratio of −0.61 and a maximum drawdown of 96%. The only profitable strategy is the unaugmented LSTM (Sharpe:0.64, CAGR:15.6%, maximum drawdown: 40%), which ranks fourth on mean MAE, holds statistically significant positive directional accuracy (51.95%, p = 0.032), remains profitable after costs of up to 30 bps and is the only strategy whose positive Sharpe ratio is statistically distinguishable from zero by block bootstrap (p = 0.008, 95% CI: [0.16, 1.14]). An inverse-MAE ensemble of the leading forecasters underperforms the LSTM on every trading metric. The findings document a pronounced and statistically verified decoupling between statistical accuracy and economic utility in emerging-market forecasting and motivate decision-aware loss functions for trading-oriented model development.

Published on 15 Jun. 2026