Prediction Market Strategy

Prediction Market Strategy

Binary prediction markets price contracts as probabilities. This skill covers the strategy, sizing, and validation layer that applies across all venues and market types. API mechanics live in kalshi-api / polymarket-api; contract semantics and settlement live in kalshi-weather-markets / kalshi-crypto-index-markets. This is the strategy/sizing/validation layer that applies across all of them.

Core Thesis

Price = implied probability. A contract priced at $0.18 claims an 18% chance of resolving YES. Brackets in a series sum to just above $1.00 — the overround is the house margin (roughly 5–8% for weather markets on Kalshi).

Takers systematically lose; makers systematically win. Across 300k+ Kalshi contracts, the average pre-fee return is ≈ −20%, concentrated in takers (market-order users) and in longshot buyers. Makers (resting limit orders) earn positive returns. On Polymarket (588M+ trades), the top ~1% of accounts capture ~76.5% of profit, predominantly by resting limit orders. This is the foundational result.

Favorite–longshot bias is the durable mechanism. Cheap longshots are systematically overpriced: a $0.05 contract historically wins ~2%; sub-$0.10 contracts lose ~60% of stake to buyers. Favorites are fairly- to slightly-underpriced. The repeatable expression is selling the overpriced longshot tail, maker-side — resting NO bids on brackets priced ~$0.05–$0.20, diversified across many events to survive the rare hit. This is structural/behavioral, not a forecasting edge.

Forecast skill ≠ trading edge. A good weather or event forecast is largely redundant with the market price at decision time. Markets aggregate information efficiently enough that even a measurably better model produces near-zero net edge after fees unless it finds systematic mispricings (which are behavioral, not informational). The exception is official-label ML in lightly-traded markets — but that is bounded by fill-rate and capacity, not forecast accuracy.

Strategy Catalog

Strategies evaluated on Kalshi/Polymarket weather and event markets. "Real" means it survived correct settlement + fees in testing; others are flagged so you don't re-chase them.

Bottom line: Two strategies survive correct accounting — the behavioral tail fade (maker-side) and the forecast-driven YES entry past the θ gate. Everything else is either HFT-scale, infrastructure-heavy, or dissolves under correct settlement + fees.

Fee-Aware Sizing & Edge Gates

All selection is on fee-adjusted net edge. Raw win-rate, return on notional, and % correct are not selection metrics.

Net Edge Formula

net_edge = p_model - ask - kalshi_fee(ask)
# Kalshi taker fee: ceil(0.07 * price * (1 - price) * 100) / 100 per contract
# Polymarket fee: 0 (no explicit taker fee; spread is the cost)

Select a trade only when net_edge > θ.

Expected-Edge Gate (θ)

Limit Price Posting

When placing maker orders, post your bid θ below model fair value:

limit_price_cents = floor((p_model - θ) × 100)

This ensures you only fill when the market moves in your favor by at least θ.

Fractional-Kelly Sizing

f_star = edge / (1 - entry_price)          # full Kelly fraction
stake = min(kelly_frac * f_star * bankroll, max_bet_fraction * bankroll)
contracts = stake / entry_price
# Defaults: kelly_frac=0.25 (quarter-Kelly), max_bet_fraction=0.015

The 1.5% bankroll cap is the binding constraint for most trades. Quarter-Kelly is aggressive enough to compound but mild enough to survive a bad run of correlated hits.

Exposure Caps (observed defaults)

Slippage Is Part of Selection

Walk the real NO/YES ladder to your full size to compute the depth-weighted entry price. If the walk pushes net_edge below θ — or slippage exceeds 50% of the edge — skip the trade. Phantom penny levels (≤2¢ asks that don't persist across snapshots and lack trade-print corroboration) must be excluded from the ladder before walking it.

Maker vs. Taker

Resting a limit order (maker) captures the spread instead of paying it, avoids (or reduces) the taker fee, and is the execution mode used by profitable accounts. The maker edge compounds the structural longshot fade. Market orders (taker) should be used only when the fill probability of a limit order is unacceptably low relative to the opportunity.

All sizing functions are in scripts/sizing.py. Run directly for a worked example.

Backtesting Methodology

Cardinal Rule: Settle on the Venue's Own Result

Never re-derive settlement from a third-party source. Use the venue's result field (Kalshi: result: "yes"/"no"; Polymarket: settlement transaction). Any derived truth that merely correlates with the venue's resolution can flip ~10% of outcomes and manufacture double-digit fake edges.

Method Checklist

• Use venue result for settlement — never self-computed truth
• Use close_time (or settlement date from the ticker), not date-of-crawl
• Decision features cut at the city/event's own local decision time — no UTC mismatch
• Entry prices from a decision-time snapshot, not near-close prices (which have already converged)
• Walk the real ladder to your size; cap by top-of-book depth; deduplicate fills
• Fee-deducted PnL only — never notional
• Temporal holdout or expanding-window walk-forward only — no shuffled CV
• Forward paper-trading before live capital: log model p, decision-time ask, and outcome; compare to the backtest Brier/accuracy