Turn probabilistic edge into compounding while surviving variance
Effective AI bankroll management starts with Kelly Criterion stake sizing, which allocates a fraction of your bankroll proportional to your edge and odds. In practice, most AI bettors use 25–50% fractional Kelly to reduce variance from calibration uncertainty. Combine this with volatility caps, exposure limits per league, and hard drawdown triggers to protect capital and enable sustainable long-term compounding.
Even a strong predictive model (Value Betting, CLV positive) can still go broke through oversized staking. Bankroll management translates probabilistic edge into controlled geometric growth by balancing three forces: expected value (EV), volatility (variance of returns) and tail risk (probability of large drawdowns).
Goal hierarchy: 1) Survival (avoid ruin), 2) Variance discipline (keep emotional + capital stability), 3) Compounding efficiency (maximize long‑run log growth).
For a single binary outcome with decimal odds O and model probability p, Kelly fraction f* = (p*O - 1) / (O - 1) assuming fair payout otherwise. Simplified for even odds: f* = 2p - 1.
Advantages: Maximizes asymptotic log utility; penalizes overconfidence. Risks: Extremely sensitive to probability miscalibration (see Calibration).
Multi-bet portfolio: Solve convex optimization with constraints Σ f_i * L_i ≤ VaR_limit where L_i is worst-case loss (stake). Approximate by summing fractional Kelly across independent bets but cap total exposure per time bucket.
Because model edge & probabilities are uncertain, practitioners scale to 0.25–0.50 Kelly (Fractional Kelly) to reduce variance drag and estimation error amplification.
Rule of thumb: If your calibration or CLV tracking shows instability (rolling 1,000 bet CLV deteriorating), cut fraction further until stability recovers.
Fractional Kelly reduces growth only modestly when uncertainty is high, while dramatically lowering maximum drawdown magnitude.
Do not allocate to every positive expected value estimate; enforce a minimum edge filter (e.g. ≥3%) like in Profit Report.
Rationale: Micro-edges near zero are most vulnerable to calibration error & fees (commission / spread) eroding EV.
Adaptive thresholding: Raise edge threshold during drawdown to concentrate risk on higher-confidence signals; relax slightly when equity at ATH (all-time high) but monitor variance metrics.
Track realized daily / weekly PnL standard deviation. If vol > target (e.g. 4% of bankroll weekly), proportionally scale all stake sizes by target_vol / realized_vol factor.
Set per-market & per-league exposure caps (e.g. max 10% aggregate risk in a single league round) to avoid hidden correlation spikes (synchronized outcomes).
In-play & pre-match mixing? Use risk buckets; pre-match commitments reduce allowable in-play Kelly fractions to preserve capital buffer.
Define hard drawdown triggers (e.g. -15%, -25%). At -15%: halve Kelly fraction; at -25%: freeze new bet types, run calibration & CLV diagnostics (see CLV Guide).
Track MAR (CAGR / Max Drawdown) and Calmar-like ratios to evaluate risk-adjusted performance; optimize for smoother equity growth, not just headline ROI.
Use rolling skew & kurtosis of returns; abrupt shift may signal model regime change (injury news modelling lag, data feed quality).
Segment capital: 1) Active betting float (hot wallet), 2) Reserve buffer (2–3 expected worst-month losses), 3) Strategic expansion fund (model R&D).
Replenish active float only after independent variance assessment; avoid emotional top-ups mid downswing without diagnostic confirmation.
Use a standardized stake ticket schema: bet_id, timestamp, league, market, raw_prob, calib_prob, edge%, stake_fraction, stake_amount, implied_vol. Enables later forensic audits.
Workflow: (1) Generate calibrated probabilities → (2) Filter by adaptive edge threshold → (3) Compute Kelly fractions → (4) Apply volatility + exposure caps → (5) Scale by fractional factor → (6) Execute & log → (7) Track CLV + calibration + drawdown dashboards.
Monthly review: Adjust fractional factor based on stability metrics (ECE, Brier delta, mean CLV).
Bankroll growth becomes a byproduct of disciplined process, not impulsive stake escalation.
Effective bankroll management converts fragile raw edge into resilient compounding. Combining calibrated probabilities (Calibration), verified market respect (CLV) and disciplined fractional Kelly produces sustainable, lower-volatility ROI.
Continue exploring foundational pillars: Value Betting for edge identification and Profit Reports for empirical performance benchmarks.
The Kelly Criterion is a mathematical formula for optimal stake sizing: f* = (p × O – 1) / (O – 1), where p is your estimated win probability and O is the decimal odds. It maximizes long-term bankroll growth by staking proportionally to your edge. However, it is highly sensitive to miscalibration — even small overestimates of edge lead to over-staking and devastating drawdowns, which is why fractional Kelly is preferred in practice.
Full Kelly assumes perfectly calibrated probabilities, which no model achieves. Even a 5% overestimate of win probability can lead to dramatically over-sized stakes and catastrophic drawdowns. Fractional Kelly — typically 25–50% of the theoretical optimal fraction — maintains most of the long-run growth benefit while dramatically reducing variance and maximum drawdown, making it the standard approach for serious AI bettors.
Drawdown triggers are predefined loss thresholds that automatically reduce betting activity. A common setup: at –15% drawdown, cut Kelly fraction by 50%; at –25%, freeze all new bets and run full diagnostic checks on calibration, CLV, and data feed quality. These rules prevent emotional over-staking during losing runs and ensure model issues are investigated before further capital is put at risk.
A well-structured betting bankroll should be divided into three segments: (1) an active float for current betting, sized based on maximum expected weekly loss; (2) a reserve buffer covering two to three expected worst-month losses to survive variance without topping up; and (3) a strategic fund for model research and account expansion. Never deploy reserve capital for active betting without a formal diagnostic review confirming the model's edge is intact.