Backtest — Simplified Model on Real NBA Data

⚠️ Honest disclaimer up front

This backtest uses a simplified model (opening spread + cumulative season win % as the only inputs). It does NOT replicate the live analysis approach we use on game pages, which includes current form, injury news, head-to-head matchup, pitcher/lineup data, and qualitative judgement.

This is the model we COULD build with what's in a free historical odds dataset. The live analyses are richer, but we have no way to backtest them retrospectively because we'd need to know what news existed BEFORE each game.

Bottom line: the result below is the test of a naive baseline, not the test of our actual approach. But the lessons it teaches are real.

📊 Headline Stats

Games considered

13,903

Bets placed

4,615

Wins

2,309

Losses

2,212

Pushes

Win rate

51.07%

Break-even at −110

52.38%

Max drawdown

96.3%

Total wagered

$27,797

Avg CLV

−4.68 pts

📉 Bankroll Trajectory

Bankroll over 4,615 bets. Starting line at $1,000. Bankroll dropped below $500 within the first 500 bets and never recovered.

🔬 What the Model Did

Data source

github.com/FinnedAI/sportsbookreview-scraper — pre-scraped NBA games 2011–2021 with opening/closing spreads, opening/closing totals, closing moneylines, per-quarter scores.

The model (intentionally simple)

For each game (chronologically):
  market_prob = Φ(−opening_spread / 11) // Gaussian conversion, σ=11
  record_signal = (home_win_pct − away_win_pct) × 0.10
  our_prob = market_prob + record_signal
  fair_market = market_prob / 1.0225 // remove half-vig
  edge = our_prob − fair_market

  IF edge ≥ 2% AND EV > 0:
    stake = min(0.02 × bankroll, 0.25 × Kelly × bankroll)
    BET on spread at −110
    Resolve via actual game margin

Sizing + gates

Edge threshold	≥ 2% over fair market
EV gate	EV per $1 stake must be positive
Sizing	Quarter Kelly, capped at 2% of bankroll
Bet odds	Standard −110 (spread market)
Vig assumption	4.5% (single-side half = 2.25% adjustment)

📈 Why It Failed — The CLV Smoking Gun

Win rate of 51.07% versus break-even of 52.38% means we lost roughly 1.3 percentage points per bet to vig. At 4,615 bets, that compounds catastrophically.

But the leading indicator is CLV (Closing Line Value):

Avg CLV

−4.68 pts

Negative CLV bets

2,161

Neutral CLV bets

812

Positive CLV bets

1,642

Translation: When we took a bet, the closing line moved AWAY from us on average. Sharp money agreed with our pick less often than they agreed with the other side. The market was processing information faster than our simple record-based model.

This is the signal of an unsharp bettor. The CLV told us within the first 500 bets that the strategy was broken — long before bankroll hit zero.

🎓 The 5 Lessons This Backtest Teaches

Lesson 1 — A simple model is a money pit. Adding "team record" as a signal to a market-derived prior produced a strategy that lost 95% of bankroll over 4,615 bets. The market already prices in records. Layering them back in just adds noise.

Lesson 2 — Win rate at 51% is meaningfully losing. 51% sounds "close to 50/50" but you need 52.4% to break even on −110 bets. The 1.4-point gap is the vig.

Lesson 3 — CLV is the leading indicator. Within ~500 bets we had clear negative CLV. We didn't need to wait for the full bankroll loss to know the strategy was broken. Track CLV in real time on live bets.

Lesson 4 — Quarter Kelly + 2% cap saved us from ruin. Even with a losing model, sizing prevented total wipeout in early bets. Bankroll ended at $49, not $0. Full Kelly would have busted within months.

Lesson 5 — Our LIVE analyses include nuance this backtest doesn't. Form trends, injuries, matchup tactics, lineup news — those aren't in the historical dataset. The simplified model failing here doesn't prove our live picks fail. But it DOES prove that simple statistical models aren't enough on their own.

📊 Per-Season Breakdown

Season	Games
2011	1,074 (lockout-shortened)
2012	1,314
2013	1,319
2014	1,311
2015	1,316
2016	1,309
2017	1,312
2018	1,312
2019	1,143 (COVID-shortened)
2020	1,171 (COVID-shortened)
2021	1,322
Total	13,903

📚 What Comes Next

The backtest validates everything the methodology page warned about:

Beating the market is hard. The market is genuinely efficient.
A 51% win rate is a LOSING strategy at standard juice.
CLV detects strategy failure faster than bankroll does.
Sizing prevents catastrophe even when the model is wrong.

For our live picks:

Track CLV on every actual bet placed
If avg CLV stays negative across 30+ live bets, the live analysis isn't beating the market either — recalibrate
If avg CLV is positive, the qualitative inputs (current form, injuries, matchup) ARE adding value above the simple model
This is the test we need to actually run with real money or paper trades over the coming months

This backtest is the floor. Our live analysis should beat it. If our live analysis just matches it, we're no better than a record-based model — and that loses money. Track CLV religiously.

📦 Reproduce This

Backtest script lives at /bets/learn/backtest-script.js.

To re-run: clone github.com/FinnedAI/sportsbookreview-scraper, set NBA_DATA env to the path of data/nba_archive_10Y.json, run with Node.js.