Probabilistic Machine Learning & Research Integrity
“We don't ask you to trust the model. We show you how we test it.”
Kingsmen Alpha uses fixed research boundaries, locked out-of-sample evaluation, execution-cost modelling, adversarial stress tests, and forward validation to reduce the risk of overfitting and self-deception. Unlike retail bots that trade incessantly, our engine operates under an anti-churn, high-hurdle discipline that defaults to cash until statistical edge clears real-world exchange friction.
LightGBM + Platt Shrinkage
Predicts directional Triple-Barrier outcomes. Post-fit Platt sigmoid shrinkage maps raw model scores into probability estimates and reduces systematic distortion, evaluated via Brier Score and Expected Calibration Error.
Top-50 FAISS Cosine Search
Compares a standardized market-state representation against historical regimes (currently 50-D) to retrieve the 50 closest analog states with Kish Effective Sample Size (N_eff ≥ 40).
Canonical 12.2 bps Friction
Deducts 8 bps fees + 3 bps slippage + 1 bp spread + 0.2 bp carry. Executes strictly if Net Expected Value ≥ +0.15%; otherwise preserves capital in cash.
Purged label boundaries with a mandatory 96-bar label purge window across dataset transitions. Features rely strictly on backward-looking data known at or before bar t; execution prices evaluated at t+1.
2024–2025 was withheld from model fitting, calibration, hurdle selection, and feature decisions. The model architecture, hyperparameters, and 30-feature Gold Matrix remained completely locked.
Subjected to worst-case same-candle TP/SL resolution (invariant SL hit recorded), 1-bar execution queue delay, and 12.2 bps round-trip friction hurdle before production eligibility.
Qualified models are sealed with SHA-256 cryptographic package hashes (Q_asset). Passing historical qualification begins an independent forward-validation clock in our 2026 sanctuary.
Machine Learning models and FAISS memory indexes are actively deployed and streaming 24/7 for all qualified assets:
Locked Out-of-Sample (OOS) Research Benchmark (2024–2025)
LOCKED OOS RESEARCH • SIMULATED| Performance Dimension | Frozen OOS Benchmark (2024–2025) | Validation Methodology |
|---|---|---|
Daily Compounded Sharpe Out-of-sample simulated back-test (2024–2025) | 4.09 Daily Sharpe¹ Simulated test of the ML model on data held back from training. 95% bootstrap CI: [3.09, 5.73]. Models 12.2 bps frictions (8 bps fees + 3 bps slippage + 1 bp spread). Hypothetical research result, not live trading. Past or simulated performance does not guarantee future results. | Stationary Block Bootstrap 95% CI: [3.09, 5.73] |
| Hypothetical Modeled Expectancy | Simulated Research Model¹ | Simulated rule-based expectancy on frozen OOS test data |
| Maximum Portfolio Drawdown | -2.52%¹ | Equal-weight 1/6 allocation across core 6 assets |
| Deflated Sharpe Ratio (DSR) | p = 0.0237 (Z = 1.98) | Corrected for 108 Recorded Research Trials |
| Round-Trip Friction Model | 12.2 bps Invariant | 8 bps fees + 3 bps slippage + 1 bp spread + 0.2 bp carry |
| Research Sample Size | 691 Research Model Samples | Strict temporal boundary separation & 96-bar purge |
¹ Sharpe 4.09 (daily, out-of-sample back-test, 2024–2025): This comes from a simulated test of the ML model on data that was held back from training. The 95% bootstrap confidence interval is 3.09 to 5.73. Modeled costs include 12.2 bps round-trip friction (8 bps taker fees + 3 bps slippage + 1 bp spread). It is a hypothetical result, not live trading. Past or simulated performance does not guarantee future results. Kingsmen Alpha does not place orders on behalf of users.
We welcome rigorous critique from quantitative researchers, economists, and engineers. If you identify theoretical improvements, mathematical inconsistencies, or edge cases in our published frameworks, contact our quantitative engineering desk directly at support@kingsmenalpha.com.