Adaptive Polynomial Regression Analysis for Predicting Indonesian Stock Exchange Movements With R-Square Optimization and Trading Channel Construction

Abstract

Stock price movements in financial markets often reflect rapid macroeconomic shifts and liquidity-driven volatility. This is particularly evident in large-cap, highly liquid stocks within the Indonesian LQ45 index. Traditional deep learning models often struggle with this market's unique volatility, suffering from overfitting, lack of transparency, and look-ahead bias. To address these issues, this study proposes an adaptive rolling polynomial regression (APR) framework for non-linear trend filtering and automated trading signal generation. The method uses a sliding window with endpoint locking to prevent look-ahead bias. At each step, it finds the best lookback window by optimizing a local OLS polynomial regression and filtering out noisy market regimes using a strict R-squared threshold. The resulting trends are converted into trading signals (BUY, SELL, HOLD, CLOSE) through dynamic Z-score volatility channels. When evaluated on daily historical data from 2015 to 2024 for five LQ45 stocks ANTM, ASII, BBCA, PTBA, and TLKM the framework shows an excellent fit, with R-squared values between 0.9881 and 0.9966. Backtesting indicates that the strategies yield positive returns and profit factors above one for all assets. ASII performed best, with a total return of 2680.16%. Despite low win rates (<10%), the long-term profitability was maintained as large trend-following gains offset small losses. This study highlights that simple, interpretable mathematical models can provide effective, computationally efficient systematic trading options in emerging equity markets.