The strategy uses LSTM2 with PCA-reduced data to predict returns for diversified ETFs, optimizing weights via Mean-Variance Optimizer, rebalancing monthly, and minimizing portfolio covariance under defined constraints.
ASSET CLASS: ETFs | REGION: Global | FREQUENCY:
Monthly | MARKET: bonds, commodities, equities, REITs | KEYWORD: Machine Learning
I. STRATEGY IN A NUTSHELL
The strategy invests in diversified ETFs across equities, bonds, real estate, and natural resources. It uses an LSTM neural network to predict returns, with inputs including macroeconomic, market, and historical asset data, reduced via PCA. Two LSTM models were tested: LSTM1 (MAE loss, 70 PCA dimensions, one hidden layer) and LSTM2 (MSE loss, 150 PCA dimensions, three hidden layers, early stopping). LSTM2 outperforms and feeds predictions into a Mean-Variance Optimizer to determine asset weights (0.05–0.35, 0 for TIPS), respecting constraints (no short selling, weights sum to 1). The portfolio is rebalanced monthly, extending the training period by one month per prediction.
II. ECONOMIC RATIONALE
Mean-Variance Optimization (MVO) seeks maximum expected return for a given risk. LSTMs, a type of recurrent neural network, use hidden and cell states with input/output gates to retain and filter past information, enabling effective return predictions. LSTM-based strategies have shown strong performance in crises (e.g., European debt crisis, China 2015 crash), though they may underperform in certain events (e.g., Taper Tantrum).
III. SOURCE PAPER
Adaptive Portfolio Asset Allocation Optimization with Deep Learning [Click to Open PDF]
Samer Obeidat; Daniel Shapiro; Mathieu Lemay; Mary Kate MacPherson; Miodrag Bolic
<Abstract>
Portfolio management is a well-known multi-factor optimization problem facing investment advisors. The system described in this work can assist in automating portfolio management, and improving risk-adjusted returns. The asset allocation action recommendations were personalized to the portfolio under consideration, and were examined empirically in this work in comparison to standard portfolio management techniques. This work presents a Long Short-Term Memory approach to adaptive asset allocation, building upon prior work on training neural networks to model causality. The neural network model discussed in this work ingests historical price data and ingests macroeconomic data and market indicators using Principal Components Analysis. The model then estimates the expected return, volatility, and correlation for the selected assets. These neural network outputs were then turned into action recommendations using a MeanVariance Optimization framework augmented to use a forwardlooking rolling window technique. Testing was performed on a dataset with a 7.66 year duration. The observed mean annualized return for classical passive portfolio management approaches were 4.67%, 3.49%, and 4.57%, with mean Sharpe ratios of 0.46, 0.20, and 0.54. 10 simulations using the new Long ShortTerm Memory model from this work provided a mean annualized return of 10.07%, with a Sharpe ratio of 0.98. This work provides the conclusion that a Long Short-Term Memory model can generate better risk-adjusted returns than conventional strategic
passive portfolio management.
IV. BACKTEST PERFORMANCE
| Annualised Return | 10.07% |
| Volatility | 10.13% |
| Beta | N/A |
| Sharpe Ratio | 0.99 |
| Sortino Ratio | N/A |
| Maximum Drawdown | N/A |
| Win Rate | N/A |
