ETH Hourly Up/Down Prediction-Market Fair-Value Gap — Crypto-Microstructure Probability Model vs Implied Price (POLYMARKET binary + BINANCE ETH signal, Long-Only Cash, Short-Horizon)

A LONG-ONLY binary strategy on the Polymarket recurring hourly ETH up/down series (ETH-updown-1h.POLYMARKET). Each interval spawns a new market whose YES token price is the market-implied probability that ETH closes the hour UP. Our structural edge: we already collect ETH spot/perp bars, so we can estimate the fair up-probability for the SAME horizon better than a spread-quoting prediction market. Mechanism (deliberately NOT a crypto momentum/mean-reversion clone — this trades a mispriced PROBABILITY, not price): from ETHUSDT.BINANCE 5-minute bars compute a short-horizon signal s = standardized last-1h log-return (momentum) and map it through a logistic p_fair = sigmoid(k * s) + d, where d is a small unconditional up-drift constant. Roughly 15 minutes before resolution, compare p_fair to the live YES implied price q. Trade only when the gap exceeds a threshold τ sized to clear the ~1% round-trip spread with margin. If p_fair − q > τ buy YES; if q − p_fair > τ buy the complement (NO) token (there is no shorting — NO is a separate token of the same series). One position per market, hold to resolution (0/1). ECONOMICS (with arithmetic): near-coin-flip markets trade around q≈0.50; entering only when the probability gap ≥ τ=0.06 gives gross EV per trade = (p_fair − q) as a fraction of the 0/1 payoff ≈ 0.06 on a ~0.50 cost ≈ +12% of notional gross, versus a ~1% round-trip spread cost — a >10x cost multiple, the opposite of the sub-fee graveyard. Binary outcomes make single-trade variance large (+100%/−100% around a 0.5 stake), so the edge only exists in AGGREGATE across many resolved markets; that is exactly what the venue supplies (a fresh market every hour). This is a genuine information edge (we see the underlying tape the market is pricing), fills the emptiest venue bucket (POLYMARKET 0.0%), is cross-venue (POLYMARKET + BINANCE, filling the 6.9%→≥15% cross-venue quota), and is short-horizon (fills 9.6% bucket). Kept to 3 parameters (momentum lookback, logistic slope k, entry threshold τ; drift d fixed) to resist the overfit graveyard.

coding_stuck_3_timeouts