Research
Deep dives into quantitative methods, model validation techniques, and the 43 S-series modules that power V7. Written to share learnings and contribute to the quant community.
How the Hurst exponent provides real-time regime classification that adapts exit giveback thresholds, improving exit timing by 18% in trending markets.
How ATR percentile-based volatility classification reduces drawdown by adjusting position sizing and signal thresholds during vol spikes.
Rolling correlation monitoring across asset clusters that detects breakdown events and prevents cascading multi-position drawdowns.
How to validate trading strategies using Monte Carlo simulation with regime-conditional resampling. Addresses the limitation of naive bootstrap that ignores market regime changes.
ADX-based momentum regime classification that identifies choppy markets and filters out low-quality signals, removing 22% of losing trades.
Quarter-Kelly position sizing integrated with DD-triggered risk zones for optimal capital allocation under FTMO constraints.
Anti-Martingale position scaling that increases risk during winning streaks and decreases during losing streaks, adding 8% to total R.
Hard daily and total loss limits with safety buffers that achieved zero FTMO breaches across 7.5 years and 4,505 trades.
ADX values as direct L1 XGBoost features, ranking in the top 5 by SHAP importance in 4 of 6 asset clusters.
Unsupervised K-Means clustering on market features achieving 0.68 silhouette score, enabling regime-conditional alpha generation.
Autoencoder reconstruction error detects regime transitions 2-3 bars before traditional indicators, providing early warning for parameter adjustment.
Rolling SHAP value analysis that tracks which features drive L1 decisions over time, detecting feature decay and concept drift.
Limits simultaneous exposure to correlated positions, reducing correlated position exposure by 40% through signal strength tiebreaking.
Proximal Policy Optimization RL agent for exit timing that improved average exit R from 0.78 to 0.92 through hold-or-exit decision optimization.
Calibrated logistic regression weighting of L1 signal confidence, where top-quintile signals achieve 67% win rate versus 52% for bottom quintile.
Sub-bar entry timing using order flow features that waits for pullbacks within signal bars, achieving 0.3 ATR better entry prices on average.
Filters trades when bid-ask spread exceeds 2x normal, eliminating 12% of trades with poor fills during illiquid periods.
Bagged ensemble of L1 models using bootstrap aggregation that reduces signal variance by 25% and provides prediction confidence intervals.
Cointegration-based pairs trading using Engle-Granger test, identifying 12 validated pairs that added 15.2R as a supplementary strategy.
Caps maximum open positions per cluster and total, preventing overexposure during high-signal periods and keeping portfolio heat at 3% max.
Realistic slippage model accounting for market impact, spread variation, and order size effects, adjusting backtest results by -2.1R for honesty.
Deep dive into Probability of Backtest Overfitting (PBO) and why it is essential for distinguishing genuine alpha from data mining artifacts.
Probability of Backtest Overfitting calculation using combinatorially symmetric cross-validation, achieving PBO=0.112 well below the 0.5 threshold.
Pre-trained LSTM on raw price sequences before fine-tuning for exit prediction, achieving 40% faster convergence through transfer learning.
Consistent feature scaling between training and inference using saved Z-score parameters, eliminating train/test preprocessing mismatch.
Anchored walk-forward optimization with expanding window confirming OOS performance within 13% of IS across 12 independent folds.
Multi-timeframe data alignment ensuring M15, H1, H4, D1 bars are properly synchronized with zero data alignment errors.
60-day rolling Sharpe ratio with decay detection alerts that flags significant deviations from historical baseline performance.
Automated trade logging capturing all 38 feature values, model confidence scores, and execution details for every trade.
M15/H1/H4 directional confluence checking that achieves 64% win rate on aligned signals versus 55% on single-timeframe.
Calendar-based filter blocking entries around major economic releases, avoiding average 1.2R loss per major news event.
Reduces position size for trades held overnight to account for gap risk, reducing overnight gap losses by 60%.
Architectural decisions behind separating signal generation (L1), entry timing (L2), and exit management (L3) in the V7 Engine.
Full walk-forward validation framework testing strategy robustness across multiple market regimes with OOS Sharpe at 87% of IS.
Validates backtest against survivorship-free instrument universe, confirming zero survivorship bias across all 4,505 trades.
Realistic commission and swap cost model per instrument accounting for spread, commission, and overnight financing at -18.3R total.
Trades the equity curve itself, reducing size when equity drops below its moving average and increasing when above.
Filters entries by trading session with session-filtered win rate of 62% versus 54% for all-hours trading.
Dynamic allocation across asset clusters based on recent momentum and regime signals, overweighting performing clusters.
Platt scaling and isotonic regression to calibrate L1 probability outputs, reducing calibration error from 8% to 2%.
High-performance data caching using Apache Parquet achieving 10x faster data loading compared to CSV.
Tracks individual feature predictive power over time, identifying 2 features with decaying alpha for potential removal.
Four-zone adaptive risk protocol scaling from 0.30% down to 0.15% based on drawdown severity, achieving 0.08% FTMO breach probability.
Block bootstrap Monte Carlo with regime-conditioned resampling validating 0.08% breach probability across 5,000+ simulations.
Calculates realistic 15% performance haircut accounting for data mining, multiple testing, and implementation shortfall.
Combinatorially Symmetric Cross-Validation testing all possible IS/OOS combinations, producing foundation for PBO=0.112 calculation.
Using Hurst exponent to detect trending vs mean-reverting regimes and adapt strategy parameters accordingly.
How LSTM networks can predict optimal exit timing by learning from trade trajectory sequences.
Every module in the S-series architecture has its own technical article explaining why it exists, how it works, and the measurable impact on performance.
Explore S-Series Architecture →