Adaptive Factor Fusion with Correlation Filtering (ACF³) for Explainable Trading Signal Detection in Technology Stocks

Abstract

Old-style trading rules like RSI, MACD, and Bollinger Bands were built once and never change, even though markets change all the time. This is why many researchers now use machine learning to find better trading signals, called alpha factors. In this paper, we build a new method called ACF³ (Adaptive Correlation-Filtered Factor Fusion). It works in three simple steps. First, it looks at all the technical indicators and removes the ones that are just repeating the same information as another indicator. Second, it tries out many different combinations of the remaining indicators and hidden (latent) patterns, and keeps only the combinations that actually predict price movement well and also perform well in a trading test. Third, it double-checks which factors really matter using two different explanation tools, SHAP and Permutation Importance, instead of relying on just one. We tested this method on six years of simulated technology-stock data made to behave like real stocks (with realistic ups and downs, and realistic correlation between companies). We used Random Forest and XGBoost as our starting models, and PCA and a small neural network (autoencoder) to find hidden patterns. On new, unseen data covering 223 trading days, our fused model correctly predicted next-day stock direction 54.3% of the time, better than the plain baseline model's 53.9%. It also caught 90.3% of the actual up-days compared to 83.2% before. When we tested it like a real trading strategy, our method earned a bit more money and had a slightly better win rate than the baseline, but it also had slightly bigger losses at its worst point. We also found that simply buying and holding the stocks the whole time did about as well as our trading strategy in this particular test period. We report this honestly instead of hiding it, because that is what real, trustworthy research should do. All the numbers, charts, and results in this paper come from code we actually ran — nothing here is guessed or made up. The data itself is a realistic simulation, not live market data, and we say this clearly throughout the paper.