Skip to content

House model: NFL Elo

Paper bot · Retired

Bets only when the model and market meet its published selection policy.

Historical record unavailable

Live record lost in the 24 Oct 2026 server move.

Simulated paper trading only — orders are not sent to Polymarket.

Model vs market picks →

The tested policy is off. The picks are still published.

The backtest found no edge worth taking, so this bot places no bets. The model still publishes and grades its picks.

Places no bets: the backtest found no edge worth taking. Picks are still published and graded.

Test ROI −38.5% across 68 test bets. The 95% interval is −61.7% to −13.9%.

All-edge baseline ROI −12.0%.

Backtest prices: closing line (nflverse), vig included. A backtest uses historical prices; the paper record above uses live quotes.

The model publishes a chance for every game and is graded on every game, whether or not this policy places a bet. A published pick is a prediction; a paper bet is an execution. This policy places no paper bets.

Every published pick and its grade →
Show policy methodology

This bot would bet on NFL game winners when its Elo ratings disagree with the price. Across 16 seasons priced at the bookmakers' closing odds nothing beat the market: the best rule found on 2020–22 (+35%) lost 38% on 2023–25, and the old 'bet every 4¢+ edge' rule lost 12% (431 bets). Betting is switched off; its win probabilities are still published and graded.

Periods
train 2010-2019 · validate 2020-2022 · test 2023-2025
Prices
closing line (nflverse), vig included
Test bets
68 · hit 29.4%
Test ROI
−38.5% (95% CI −61.7% to −13.9%)
Every ≥4¢ edge
−12.0%
Version
2026-10-02

In plain words

TL;DR

Off: the Elo model publishes picks on NFL moneylines, but the bot places no bets because the backtest found no edge after fees.

How it bets

Every team has a rating that rises after wins and falls after losses (more for big margins, adjusted for home field). The rating gap becomes a win chance; then see "Why 4¢" below.

Why this bot is off

  1. Before each game the model publishes its chance for every side, next to the market's price.
  2. In the backtest, betting these picks lost money once the bookmaker's or Polymarket's cut is paid.
  3. So the bot sits out. Not betting is a result too: the market priced these games better than the model.

Exact rules

Off: the backtest found no edge worth betting, so this bot places no bets. The model still publishes a pick for every game so its record stays public. This bot would bet on NFL game winners when its Elo ratings disagree with the price. Across 16 seasons priced at the bookmakers' closing odds nothing beat the market: the best rule found on 2020–22 (+35%) lost 38% on 2023–25, and the old 'bet every 4¢+ edge' rule lost 12% (431 bets). Betting is switched off; its win probabilities are still published and graded.

Bets

No published bets yet: this record has no settled, closed or open bets.