Our tracked record bets a flat 1 unit on every pick. Not Kelly. Not tiered. One unit, every time. This page explains why — and shows the Kelly numbers as a reference framework, clearly labeled as theory, not advice.
Every pick in our tracked ledger is graded at a flat 1 unit, regardless of tier. An S-tier pick and a C-tier pick stake the same on the record. We do this because it is the only staking policy that survived backtesting, and because it is the only honest way to report results: when you read “up 12 units,” that number is not inflated by a sizing scheme that quietly leveraged the winners.
A unit is whatever you decide it is — most people use 1% of a dedicated bankroll. The point of a unit is that it scales with you and keeps performance comparable across bankroll sizes. What matters here is that we never size up on “locks.” There are no locks.
The Kelly Criterion gives the bankroll fraction that maximizes long-run growth. On paper it is optimal. In practice it assumes your win probabilities are exactly right. Ours are not exactly right — no model’s are — and Kelly punishes that error brutally.
We ran it. On this model’s disagree strategy, full Kelly comes out to a wager of roughly 21% of bankroll per bet, with a ~99% maximum drawdown across the backtest. That is not a typo. A staking plan that can take you down 99% from your peak is not a staking plan; it is a coin flip with extra steps. Fractional Kelly softens it but the variance is still unacceptable for a 6-season, ~42-bets-per-year sample.
Flat betting wins not because it grows fastest, but because it survives. The model’s real edge is roughly 10.8 points over the break-even win rate, and a strategy only collects that edge if it is still solvent when the edge shows up. Flat staking keeps you in the game through the inevitable cold stretch — the longest losing streak in the backtest was 5 — without turning normal variance into a wipeout.
We still compute a Kelly number per pick, because it is a useful way to compare the relative edge between tiers, and you will see a small K Xu badge on each pick card. Treat it the way you would a speedometer in a parking lot: informative, not an instruction. The table below shows each tier’s cap alongside a theoretical quarter-Kelly stake at representative odds (−110 and +141, the average price on our disagree bets).
| TIER | CONVICTION | CAP | ¼-KELLY (THEORY) |
|---|---|---|---|
| S | Highest conviction | 3.0u | 2.0–3.0u |
| A+ | Strong edge | 2.0u | 2.0u |
| A | Solid edge | 1.5u | 1.5u |
| B | Modest edge | 1.0u | 0.5–1.0u |
| C | Lean only | 0.5u | 0.0–0.5u |
| TRACKED | Every pick, every tier | — | 1.0u flat |
Notice the gap between the cap and what we actually stake. The cap is the ceiling Kelly is allowed to suggest. The flat 1u line is what we bet. We publish both so nothing is hidden — but if you take one number from this page, take 1u.
If you bet our picks, here is the framework that matches how the record was actually built:
The instinct to size up on the S-tier “locks” is the instinct that blows up bankrolls. We built the entire record on the discipline of not doing that. If a quarter-Kelly framework helps you reason about relative edge, use it as a lens. Do not use it as a staking plan.
The record you see on the results page is built on flat staking. No leverage, no hindsight sizing, no “we’d have crushed it with Kelly.”