portfolio-optimization

This paper investigates the integration ofdeep learning-based volatility forecasting with portfoliooptimization strategies. We develop and evaluate aframework that combines three neural architectures\u2014ResNet1D, WaveletCNN, and Temporal ConvolutionalAutoencoder\u2014with both classical mean-variance optimization and reinforcement learning approaches. Usinga comprehensive dataset spanning 2012-2025, we systematically analyze how different volatility-sentiment indicators (DIX, GEX, PCR, SKEW, VIX) and rebalancingfrequencies affect portfolio performance across eightlarge-cap stocks. Our findings reveal that WaveletCNNcoupled with reinforcement learning achieves superiorraw returns (23.5% annually), while Markowitz optimization delivers more consistent risk-adjusted performance (\u03c3Sharpe = 0.03 vs. \u03c3Sharpe = 0.15 for RL). We identify clear temporal dynamics in strategy effectiveness,with RL performing optimally at weekly rebalancingintervals versus quarterly for Markowitz. Furthermore,multi-factor combinations systematically outperformsingle-factor approaches, with the DIX+GEX+SKEWensemble yielding a 26.6% annual return under semiannual rebalancing. The results demonstrate a fundamental trade-off between return maximization and outcomeconsistency, with RL strategies generating positivelyskewed return distributions (skewness = 1.43) versusMarkowitz\u2019s more compact, near-normal distribution.This research contributes to the growing intersectionof machine learning and quantitative finance, offeringpractitioners evidence-based guidance for selectingvolatility forecasting architectures and optimizationstrategies aligned with specific investment mandatesand risk preferences.