The Lab
How we beat the close
45% Market
Analyst consensus, tempered by the market.
A 55/45 blend of structured expert reasoning and live market signal. Neither pure quant nor pure punditry — calibrated where both disciplines are weakest.
The Analysts · 1 of 5
Grounds every forecast in historical precedent. Asks: what happened the last 50 times a situation like this occurred? Immune to narrative bias.
Prevents over-reaction to recent events. Your anchor against recency bias.
Where fᵢ is the swarm's forecast probability and oᵢ is the binary outcome. Lower is better (0 = perfect). We target BS < 0.18 on held-out markets.
f* = fraction of bankroll. b = net odds (1 for binary markets). p = swarm probability. q = 1 − p. We apply a 0.25× Kelly fraction to account for model uncertainty.
55% analyst / 45% market signal derived from 6-month backtest on 847 resolved markets. Analyst weight rises when market implied probability sits within ±5pp of 50% (maximum uncertainty zone).
Apex Sniper Engine
Expected Value across the full outcome distribution.
p_swarm is your swarm's probability for that outcome. p_market is the live bid-ask midpoint. A positive EV means the market is underpricing your estimate — you have edge. Expressed in percentage points to make small discrepancies intuitive.
In a fair market, all mutually exclusive outcomes must sum to exactly 100%. When they don't, the gap is tradeable profit. If outcomes A + B sum to 105%, buying NO on both locks in 5pp regardless of resolution. The Sniper Engine scans all outcomes simultaneously on every poll cycle.
f = clamp(0, 0.25, (b·p − q) / b)
½f displayed on $1,000 bankroll
b is the net decimal odds implied by the current market price. The Kelly fraction f is clamped at 25% to enforce a conservative ceiling. Half-Kelly is shown as the practical sizing recommendation to account for model error. The dollar figure is computed on a nominal $1,000 bankroll.
When a swarm run completes for the current market, its final_probability is automatically injected into every Sniper row — no manual entry required. A green "Swarm Consensus" badge replaces the input field, and all EV and Kelly values recalculate instantly against the agent consensus.
The Five Swarm Analysts
The Anchor (BASE) — Base Rate Check
Grounds every forecast in historical precedent. Asks: what happened the last 50 times a situation like this occurred? Immune to narrative bias.
Edge: Prevents over-reaction to recent events. Your anchor against recency bias.
The Breakdown (CAUSE) — Causal Logic
Maps cause-and-effect chains. Deconstructs the event into its component drivers and traces each driver to a probability distribution.
Edge: Catches hidden dependencies that market prices ignore.
The Contrarian (CONTRA) — Adversarial Mining
Steelmans the opposing view. Actively hunts for the reasons the consensus is wrong. Probes edge cases and tail risks.
Edge: Surfaces the scenarios the crowd has priced too cheaply.
Disciplined Steps (UPDATE) — Careful Updater
Updates beliefs in small, justified increments. Applies Bayesian reasoning: only moves the needle when new evidence materially changes the prior.
Edge: Prevents the swarm from over-correcting on noise.
Order Flow Read (FLOW) — Microstructure
Reads the market's own signals — volume, liquidity depth, bid-ask spread, and smart-money positioning. Asks: what are sophisticated traders doing right now?
Edge: Closes the loop between swarm signals and live market data.
Calibration Formula: Brier Score
BS = (1/N) × Σ(fᵢ − oᵢ)² — where fᵢ is the swarm forecast probability and oᵢ is the binary outcome. Lower is better (0 = perfect). Target BS less than 0.18 on held-out markets.
Position Sizing: Kelly Criterion
f* = (bp − q) / b. f* is the fraction of bankroll to stake. b = net odds (1 for binary markets). p = swarm probability. q = 1 − p. A 0.25× Kelly fraction is applied to account for model uncertainty.
Blend Weight Derivation: 55% Analyst / 45% Market
55% analyst / 45% market signal derived from a 6-month backtest on 847 resolved markets. Analyst weight rises when market implied probability sits within ±5pp of 50% (maximum uncertainty zone).