# Methodology — FightIQ Bot Backtest

Last updated: 2026-04-25 13:33Z

## What this is

We ran the FightIQ model picks against every UFC fight that had a liquid Polymarket
market over a 7-month window (June 2025 to January 2026), using the same decision
rules the live bot uses today. Every single trade is published in
`bet_ledger_public.csv` — no cherry-picking, including the five losses.

## What "walk-forward" means

The FightIQ model is trained once at the start of each year on data from BEFORE
that year. It then predicts every fight in that year without ever seeing future
data. There is no look-ahead leakage. A 2025 prediction was generated from a
model that only knew 2024 and earlier.

## How we built the trade ledger

1. Downloaded every Polymarket UFC trade since February 2023 from their public
   Gamma API (~250,000 trades).
2. Filtered to markets with at least $5,000 of cumulative volume — illiquid
   markets have wide spreads that would eat any edge.
3. For each fight, computed:
   - Opening price = volume-weighted average of the first 6 hours of real
     trading (excluding the 2-day cold-start period after listing).
   - Closing price = volume-weighted average of trades from T-6h to T-1h
     before the fight.
   - We deliberately exclude the final hour because Polymarket markets close
     several hours AFTER the fight, so prices in the last hour reflect
     near-resolved outcomes, not pre-fight betting.
4. For each strategy's trigger condition (see `strategies.json`), we logged
   whether it fired, what side it bet, and at what price.
5. We compared bet outcome to actual fight result (winner side, method, end round).
6. Stakes are flat $200 per bet across the entire backtest. We do NOT compound.
   A compounding model on this win rate runs the math to absurd numbers
   (>$200M from $10k) which doesn't reflect reality — Polymarket prop markets
   only have $1k-5k of book depth, so large bets can't actually be filled.

## What we found

Across 93 trades over 7 months (averaged about 13 per month):

- 88/93 wins (94.6%)
- Total P&L on flat $200 stakes: $45,462
- Average P&L per trade: $488.84
- Per-trade ROI: 244.4%

The single biggest contributor is Strategy 6 (Sell "goes the distance" YES) —
it accounts for less than a quarter of trades but the majority of dollar P&L.
That strategy is also the one with the smallest sample (n=22) and is most
likely to degrade out-of-sample. The website flags this clearly.

## What we deliberately do NOT publish

- Model internals (features, weights, training code)
- Live wallet address or private keys
- Real-time signal feed (we don't want copy-trading bots eating our edge)
- The live bot's exact trigger thresholds when they differ from these
  publishable summaries (we may tighten thresholds for live trading)

## What we DO publish

- Every backtest trade, including all 5 losses
- The 4 strategy descriptions in plain English
- Win rates, P&L, and confidence intervals
- Sample sizes and the small-sample caveats
- Monthly results — every month, including the bad ones

## Important caveats

1. **Backtest only.** Bot is in dry-run mode for Vegas 116 (Apr 25). No real
   money has changed hands.
2. **Filter selects for liquid markets.** Out-of-sample, expect win rates to
   drift to 80-90% from the in-sample 94.6%.
3. **Strategy 6 is the whale.** Drives most of the dollar P&L from a tiny
   sample. If you remove it entirely, the remaining 3 strategies still return
   roughly +400% on $10k bankroll over 7 months.
4. **Edges may erode.** As Polymarket UFC volume grows and sharper traders
   arrive, the structural mispricings we exploit could tighten. We monitor
   monthly and re-validate before each new card.
5. **Past performance does not guarantee future results.**