Prediction-Driven Downside-Risk Portfolio Optimization Using Forecast Error-Based Mean Semi-Absolute Deviation

Abstract

The paper proposes a prediction-driven portfolio framework that explicitly models downside risk arising from systematic over-prediction in return forecasts, departing from conventional predict-then-optimize methods that rely on historical return variability. A new downside forecast error (DSFE) risk measure is embedded into a mean semi-absolute deviation (MSAD) portfolio optimization. Three deep learning models—DMLP, LSTM, and CNN—forecast future OHLCV values, from which return forecasts are derived. Forecasts and portfolio weights are generated within a walk-forward rolling retraining scheme to avoid look-ahead bias. Portfolio risk is defined as the MSAD of portfolio-level downside forecast errors (losses from overly optimistic predictions). Performance is evaluated net of transaction costs and benchmarked against equal-weight, mean–variance, MAD, and CVaR portfolios across multiple risk-adjusted metrics. Using Abu Dhabi Securities Exchange data, Diebold–Mariano tests find no statistically significant difference in predictive accuracy across the three architectures. At the portfolio level, however, the LSTM-based DSFE-MSAD implementation performs best among DSFE-MSAD variants—highest annualized return, Sortino ratio, and terminal wealth. The DSFE-MSAD approach does not uniformly beat all benchmarks: the CVaR portfolio outperforms it on risk-adjusted and terminal-wealth measures in the main one-day-ahead test. The contribution is thus framed as prediction-consistent downside risk integration, not universal outperformance. The framework offers an implementable tool linking predictive uncertainty to downside risk for managers in volatile/emerging markets; transaction costs and turnover are shown to materially affect realized performance, especially for forecast-responsive strategies like DSFE-MSAD. The study introduces a downside-risk measure based on portfolio-level forecast errors, integrated with MSAD optimization under walk-forward, cost-aware conditions—bridging deep-learning forecasting with downside-risk portfolio theory.