The Parlay Logic AI Hedge Math Methodology

The first rule of the methodology is a data-integrity rule rather than a mathematical one. When a wager is saved, its stake and price are frozen. Live market movement updates the protection side of the calculation only. If a displayed original price ever changed to match a live price, every downstream number would be wrong, and the user's actual financial position would be misrepresented.

The live opposite ladder is not a single price. It is every available line on the opposing side — alternate spreads, alternate totals, the moneyline — each with its own price and its own settlement behaviour. Evaluating only the mirror line would miss most qualifying protection, and would also miss most push traps.

A protection pair is only real if the two wagers are settlement complements across every outcome that can occur. Line shopping across an alternate ladder makes this non-obvious: a candidate can look like a mirror while quietly leaving a gap on one specific final margin.

Because the grid is enumerated rather than inferred, the engine cannot be fooled by a half-point that looks harmless. A -3.5 original paired against a -3 opposite fails on a three-point margin; a -3 original paired against +3 fails on the same margin for a different reason. Both are rejected before pricing is considered at all.

Each mode is the same equation with a different target. Solve for the protection stake Sₕ that produces the desired result on the weaker covered outcome, then verify the stronger outcome remains acceptable.

The equalising stake is Sₕ = R₀ ÷ Dₕ. Every other target is a scaled version of that. For Break Even, solve Sₕ × Dₕ − Sₕ − S₀ ≈ 0, which reduces to Sₕ ≈ S₀ ÷ (Dₕ − 1). For a Guaranteed Profit check, verify that both min(R₀ − S₀ − Sₕ, Sₕ × Dₕ − Sₕ − S₀) is positive and clears the 20% floor.

The cap exists because protection stops being protection at some multiple of the original position. A $100 wager protected by a $340 hedge is not a protected $100 wager; it is a $340 bet with a $100 offset. Users consistently underestimate how quickly required stakes grow as the opposite price shortens.

The cap rarely binds on moneyline protection and binds frequently on short alternate lines, which is exactly where naive tools produce their worst recommendations. When the cap binds, the engine reports OUT_OF_BAND rather than silently offering a partial stake that no longer achieves the stated target.

This ordering is deliberately defensive. A candidate with a spectacular best case and a marginal worst case is worse protection than one with a slightly lower ceiling and a materially higher floor, because the entire purpose of the exercise is the floor.

1. Worst-case net — the least favourable covered settlement outcome, maximised.
2. Lowest stake — among comparable floors, the cheaper protection wins.
3. Lowest total exposure — less money on the table at equal benefit.
4. Return on risk — used only to break remaining ties.

An engine that always returns a hedge is not measuring anything. Across a full slate, most tracked positions spend most of their time in Hold or Watch, because live prices are usually efficient relative to game state. Publishing that plainly is a correctness requirement, not a UX compromise.

There is one additional rule layered on top: a no-vig implied win probability at or above 80% on the original position returns Hold regardless of what the ladder offers, unless a qualifying Guaranteed Profit candidate overrides it. Protecting a position that is already overwhelmingly likely to win almost always destroys more value than it preserves.

Computation and notification are separate concerns. Every tracked position is evaluated repeatedly whether or not anything is sent. Notification requires both a genuine state transition and a material change, plus a cooldown to prevent a flapping market from producing a stream of contradictory messages.

Downgrades are additionally damped. A single stale or unpriced cycle does not flip a recommendation to Unavailable; the condition must persist across consecutive evaluations. This prevents a momentary market suspension from being reported as a lost opportunity.

Published math is only meaningful if the outputs can be checked. Each evaluation stores the frozen original terms, the ladder snapshot used, the settlement grid results, the ranking order, and the final state. When a settlement is later voided or regraded, the correction is recorded as new evidence rather than overwriting the original record.

If a published example on this site is found to contain an arithmetic error, it is corrected and the correction is disclosed under the editorial policy. The formulas on this page are the reference; anywhere an app display disagrees with them, treat this page as authoritative and report the discrepancy.