Systematic Edge in Undermodeled Markets

APOLLO V9.3 — Deploying $20K on Czech Liga Pro Table Tennis

+2.71%  ROI Backtest OOS
1.87  Sharpe Backtest OOS
4,204  Bets Validated
See Live Results ↓

Why Czech Liga Pro?

01

Structural Inefficiency

Czech Liga Pro is a high-frequency table tennis league operating 12+ hours daily with minimal bookmaker attention. Market-makers rely on generic algorithms, creating persistent mispricings in player-specific fatigue dynamics that sophisticated models can exploit.

02

Unpriced Signal

Physical fatigue is the dominant unpriced factor in continuous table tennis leagues. Players competing in 6+ matches per day exhibit measurable performance degradation that bookmaker odds fail to incorporate. Our residual model captures this systematic error.

03

Defensible Moat

Replication requires 113K+ match historical dataset, proprietary feature engineering pipeline, domain-specific Elo tracking, and real-time state management infrastructure. The data collection alone represents 18+ months of continuous ingestion from a specialized API vendor.

Real-Time Trading Results

Win Rate
53.1%
Max Drawdown
$5,127.38 (18.93%)
Cumulative PnL
+$4,304.38
Total Bets
553
Avg Edge
Equity Curve +$4,304.38
Trade Tape 0 trades
Time Asset Signal Odds Edge% Stake PnL Result

Note: This is a live tracking period with real capital. All metrics update in real-time as bets are placed and resolved. Bets marked CASHOUT reflect early position exits with actual P&L (payout minus wager), not model-driven resolutions.

The Engine

01

CatBoost Ensemble

Gradient-boosted decision trees trained on 69K+ historical matches. Residual architecture predicts bookmaker error, not raw outcome.

02

Isotonic Calibration

5-fold cross-validated probability calibration. Maps model scores to true event frequencies. Brier CV: 0.1496.

03

Kelly Sizing

Fractional Kelly criterion with Brier-scaled dynamic brake. Auto-deleverages when model calibration drifts beyond tolerance.

Metric Value
ROI / Bet (OOS)+2.71%
Sharpe Ratio (annualized)1.87
Max Drawdown3.30%
Feature Families5 (8 residual signals)
Total Bets (OOS)4,204
Test-Match Universe36,908
2σ Sharpe Hurdle1.92

Signal Contribution by Family

Signal α
~15%
Signal β
~33%
Signal γ
~14%
Signal δ
~23%
Signal ε
~15%
Request decoded mapping →

Signals α through ε represent 5 proprietary feature families covering Fatigue & Recovery, Skill & Rating, Form & Momentum, Market Pricing, and Contextual dimensions. Exact family-to-signal mappings are withheld; qualified investors may request the decoded breakdown under NDA.

Rigorous Out-of-Sample Testing

Every claim is backed by institutional-grade validation. We use techniques from Marcos López de Prado’s Advances in Financial Machine Learning to prevent overfitting and ensure genuine predictive power.

CPCV

Combinatorially Purged Cross-Validation

Eliminates information leakage between train/test folds by purging overlapping temporal data. Standard k-fold is invalid for time-series betting data — CPCV is the gold standard.

Walk-Forward

4,204 OOS Bets · 36,908 Test Matches

Sequential out-of-sample testing that mimics real deployment. The V9.3 model is trained on chronologically prior data and evaluated on a strictly future window with point-in-time feature state.

Bootstrap

Sharpe CI₉₅: [−0.29, +4.00]

10,000-iteration block bootstrap on the OOS bet log. CI₉₅ is reported honestly: V9.3 has a positive Sharpe point estimate (+1.87) but the lower bound straddles zero, consistent with a thin-edge system requiring continued sample accumulation.

Clean Out-of-Sample Backtest

Metric Value
Total Bets (OOS)4,204
ROI per Bet+2.71%
Sharpe Ratio (annualized)1.87
Sharpe 95% CI[−0.29, +4.00]
Max Drawdown3.30%
Live Model Brier0.2403
Live Bookmaker Brier0.2423

Backtest results from V9.3 walk-forward evaluation (36,908 test matches, 4,204 qualifying bets) with no look-ahead bias. Sharpe CI₉₅ reported honestly: V9.3 is a directionally positive, statistically thin-edge system. Live shadow-trader Brier (0.2403 vs bookmaker 0.2423) confirms the model is marginally better calibrated than the market on resolved live bets.

Defense in Depth

JIT Execution Firewall

Matches evaluated at T-30min with finalized player state. Eliminates phantom rest artifacts from early API polling.

Event Sourcing

Bankroll reconstructed deterministically from first principles: $20,000 + SUM(PnL) - SUM(sweeps). Immune to state corruption.

Capital Sweep

Profits above the high-water mark swept to reserve ledger, resetting operating bankroll. Locks in realized gains.

Forecast Ledger

Matches >30min out logged to market_projections for research. Re-evaluated with fresh state when entering JIT window.

Disclaimer: Past performance does not guarantee future results. All trading and wagering involves risk of loss. This system deploys real capital and results shown reflect actual positions taken. Not investment advice.