Kairos — Systematic Equity Strategy (Machine Learning)
Conclusion
10-year walk-forward: +27.8% CAGR, 1.16 Sharpe, +12.7% alpha vs the S&P 500
- Walk-forward CAGR
- +27.8%
- Sharpe ratio
- 1.16
- Alpha vs S&P 500
- +12.7%
- Max drawdown
- 18.5%
- Vol forecaster
- LightGBM, R² ≈ 0.77
- Timing skill
- T–M γ +3.78, p<0.001
Kairos is a systematic US large-cap equity strategy built end-to-end in Python that pairs a LightGBM volatility forecaster with a Soft Actor-Critic reinforcement-learning allocator and integrated risk management across roughly 50 US large-caps. The forecaster reaches an R² of about 0.77 (Spearman ρ ≈ 0.81), feeding a learned allocation policy rather than a static optimizer. Over a 10-year walk-forward evaluation the strategy delivered a +27.8% CAGR, 1.16 Sharpe, +12.7% alpha versus the S&P 500 and an 18.5% maximum drawdown, with performance stress-tested through formal factor and market-timing regressions.
Forecasting and allocation architecture
The engine separates prediction from decision-making. A gradient-boosted LightGBM model forecasts forward volatility, achieving an R² of roughly 0.77 and a Spearman rank correlation near 0.81 — strong enough that its ranking signal can drive position sizing. A Soft Actor-Critic reinforcement-learning agent then acts as the portfolio allocator, learning an allocation policy over the roughly 50-name large-cap universe rather than solving a single-period mean-variance problem. Integrated risk management sits alongside the allocator, so exposure is governed continuously rather than reset only at rebalancing.
Out-of-sample validation and attribution
10-year walk-forward headline metrics
Annualized return, alpha vs the S&P 500, and maximum peak-to-trough drawdown
Out-of-sample walk-forward evaluation; drawdown shown as magnitude.
Results are reported on a 10-year walk-forward basis to keep the evaluation honest and out-of-sample: +27.8% CAGR, 1.16 Sharpe, +12.7% alpha against the S&P 500 and an 18.5% maximum drawdown. Under Fama–French factor controls the strategy retains statistically significant alpha (regression R² = 0.19), indicating returns are not merely factor beta. Market-timing tests are significant as well — a Treynor–Mazuy γ of +3.78 (p < 0.001) and Henriksson–Merton downside protection (p = 0.02) — evidence of genuine timing skill rather than a single lucky regime.
Downside behavior across bad markets
Across the 10 worst market months in the evaluation window, Kairos posted positive returns in all of them, consistent with the significant downside-protection coefficient from the Henriksson–Merton test. Combined with the contained 18.5% maximum drawdown against equity-like upside, this profile reflects the intended design: the volatility forecaster and reinforcement-learning allocator jointly de-risk into stress rather than riding the market down.
The Mantike engine and 20-year cycle
Mantike vs SPY over the 20-year cycle
Higher compounding with less than half the drawdown (drawdown shown as magnitude)
20-year walk-forward simulation, 2006–2026.
Kairos is complemented by the Mantike engine, a separate 20-year walk-forward simulation using a volatility forecaster (Spearman ρ ≈ 0.81) feeding rolling retraining, Markowitz allocation and ridge-based regime detection. Over the full 20-year cycle it simulated a 25.6% CAGR versus 10.3% for SPY, a 1.17 Sharpe and a −23.3% maximum drawdown against SPY's −55.2%. Together the two engines demonstrate the same thesis across horizons: pairing machine-learned volatility forecasts with disciplined allocation and regime awareness can compound well above the benchmark while materially cutting drawdowns.