Kelly gives a stake from a price and a probability. It does not know whether the probability is fragile, whether several bets fail together, or whether losing the suggested amount would change your life. A loss-floor policy puts those human constraints back into the decision before the ticket is placed.
The name is awkward; the control is simple
What many bettors call a loss floor is functionally a loss ceiling. The allowed stake is the smaller of the model-derived stake and a cap chosen outside the model:
allowed stake = min(model stake, loss ceiling)
The formula is intentionally boring. The hard work is defining the bankroll, deciding which losses count together, and setting a ceiling that remains binding when the bettor is excited, angry, or certain the next position is different.
The underlying Kelly mechanics are covered in the Kelly criterion guide. The cap is not an improvement to the probability model. It is a boundary around what the operator permits the model to risk.
Why the cap must be independent of conviction
Kelly increases the suggested stake when the estimated edge increases. Unfortunately, the largest estimated edges are often where the assumptions are least familiar: a new market, a late injury, a sparse prop, or a model operating outside its training distribution.
If the same confidence score controls both the model stake and the risk cap, the control disappears exactly when it is needed. The ceiling must come from bankroll policy, not from the strength of the current opinion. It should be written before the slate is seen and changed only through a documented review.
A cap is especially important when the probability cannot be calibrated against a real ATS record with wins, losses, win rate, window, and graded sample size. No history means no evidence for aggressive sizing.
Define the bankroll before defining the stake
The bankroll is a segregated pool that can absorb the permitted loss without touching ordinary obligations. Credit, emergency savings, and money reserved for bills are not bankroll. Depositing more after a losing session does not retroactively make the original exposure sound.
Bankroll management basics provides the accounting frame. The loss ceiling should be expressed against that defined pool and recomputed only when the bankroll changes under the written policy.
Without a stable denominator, a percentage cap is theater. The operator can make any stake appear small by silently redefining the bankroll.
Per-position limits do not solve correlated exposure
A team side, a game total, and several player props can share one assumption. Each ticket may sit below the per-position ceiling while the combined slate carries a much larger loss if the assumed game script fails.
Group positions by common driver: game, team, player availability, weather, model version, or another dependency that can break together. Apply an aggregate ceiling to the group. When the dependency is unclear, classify the bets as correlated rather than granting independence by default.
This is also where parlays and same-game combinations become dangerous. Packaging correlated outcomes into one ticket does not make the risk disappear. It concentrates it.
Add a session stop that cannot negotiate
The per-position ceiling limits one decision. A session stop limits a sequence of decisions. Once realized and defined open losses reach the stop condition, new exposure is blocked until the policy’s reset time or review event.
The implementation described in the client-side loss-limit guide should be treated as a boundary, not a warning banner. A control that can be dismissed during a losing run is a suggestion.
Write the failure state explicitly: limit reached; new wager rejected; cure is to wait for the defined reset and review the ledger. Do not cure it by changing markets, switching accounts, or calling the next ticket a hedge.
Model uncertainty belongs in the rule
A point estimate hides a range of plausible probabilities. Before sizing, ask what the stake would look like at the less favorable end of that range. If the edge disappears under a reasonable error allowance, the honest classification is low confidence or pass.
Fractional Kelly can reduce sensitivity, but it should sit inside the ceiling rather than replace it. The policy becomes layered: validate the market and probability, compute a model stake, scale for uncertainty, apply position and group ceilings, then apply the session stop.
Each layer can return an error value with a cure: uncalibrated model, verify the forecast; mismatched market, rebuild the target; exposure cap reached, reduce or reject; session stop reached, block new action.
Review the policy on schedule, not after pain
Risk rules should change through a fixed review with the ledger open. Use real settled rows and disclose the window and sample size. Do not raise the ceiling because a short run won, or lower it as punishment immediately after a loss. Both reactions let outcomes rewrite a policy that was supposed to control outcome-driven behavior.
The desk can surface current exposure, but the governing rule should remain legible outside the interface. Another person should be able to read the policy and determine whether a proposed stake is allowed.
The loss floor is not a promise of safety. It is a refusal to let one confident estimate write an unlimited check. That is a narrower claim, and a more useful one.
