Key Idea: What Is This Paper About?
Avramov, Cheng, and Metzker rigorously test two popular deep learning strategies (GKX and CPZ) under standard economic restrictions like excluding microcaps, accounting for distress, and turnover costs. They find that most of the alpha disappears once these frictions are applied. However, machine learning models still perform well on the long side, during crisis periods, and tend to pick mispriced stocks consistent with known anomalies.
Economic Rationale: Why Should This Work?
📌 Relevant Economic Theories and Justifications:
- Limits to Arbitrage: Alpha is concentrated in hard-to-trade stocks—microcaps, illiquids, distressed names.
- Arbitrage Frictions: Short-sale constraints, turnover costs, and extreme positions kill real-world performance.
- Market Inefficiency During Stress: ML strategies shine in high-volatility, low-liquidity environments (e.g., 2008).
- Anomaly Alignment: ML signals mimic combinations of value, momentum, profitability, and investment characteristics.
📌 Why It Matters:
ML models aren’t magic—they're subject to the same frictions as other models. Evaluating their robustness to economic filters is key to real-world deployment.
How to Do It: Data, Model, and Strategy Implementation
Data Used
- Period: 1987–2017 (GKX), 1967–2016 (CPZ)
- Stocks: CRSP/Compustat firms (~22,000+ over full period)
- Characteristics: 94 (GKX) and 46 (CPZ) firm-level signals + macro variables
- Training/Testing: 18-year training, rolling out-of-sample window
Models Compared
- GKX (Gu, Kelly, Xiu): NN3 feedforward deep net
- CPZ (Chen, Pelger, Zhu): GAN-based no-arbitrage model estimating SDF
- KNS (Kozak, Nagel, Santosh): Ridge regression of MVE portfolio on factors
Economic Restrictions Applied
- Subsamples:
- Exclude microcaps (NYSE < 20th percentile)
- Exclude unrated or distressed firms (around downgrades)
- Turnover Constraints: Tracked monthly (~87%–168%)
- Position Constraints: CPZ/KNS implied portfolios take +/−150–200% positions
Key Table or Figure from the Paper
📊 Reference: [Table 5] – Downside Risk and Turnover of Machine Learning Portfolios
📌 Explanation:
This table evaluates the risk and cost profile of machine learning-based long-short portfolios, focusing on Sharpe ratio, skewness, kurtosis, drawdowns, crisis performance, and turnover. It compares NN3 (GKX), CPZ (GAN-based SDF), and the Market Portfolio under full and restricted samples.
🔍 Key Highlights:
-
Sharpe Ratios (Annualized, Value-Weighted):
- GKX: 0.94 (full), drops to 0.48 (excluding microcaps)
- CPZ: 1.23 (full), drops to 0.56 (excluding microcaps)
- Market: 0.53
-
Downside Risk:
- GKX/CPZ portfolios have positive skewness and lower drawdowns than market
- GKX drawdown: 35% vs. 49% for market
- CPZ drawdown: only 21%
-
Crisis Performance:
- GKX: +2.93% to +4.1% per month during major crises (e.g., 2008)
- CPZ: −0.02% to +0.9%
- Market: −6.91%
-
Turnover (Monthly):
- GKX: ~87–99%
- CPZ: ~162–168%
- Transaction Costs Implication:
- Each 50% turnover → ~0.5% drag on returns
- Adjusted alpha becomes negligible in filtered samples
📌 Conclusion:
While ML signals show strong returns and risk resilience in-sample, their high turnover and fragility under economic constraints (e.g., microcap exclusion) raise concerns about real-world implementability.
Final Thought
💡 ML can find alpha—but you better check if it survives the real world. ⚙️📉
Paper Details (For Further Reading)
- Title: Machine Learning versus Economic Restrictions: Evidence from Stock Return Predictability
- Authors: Doron Avramov, Si Cheng, Lior Metzker
- Publication Year: 2019
- Journal/Source: SSRN Working Paper
- Link: https://ssrn.com/abstract=3345183