I Used AI to Predict 24 NBA Games: Starting With a 0% Exact-Score Hit Rate

Drafted: 2026-05-06 · Published: 2026-05-12
Suggested category: Retrospective / Sports (not the default Web3)

The Origin: Setting the Tone in One Sentence

Over the past month I've treated walker-learn.xyz/predictions/ as a public AI prediction sandbox—posting predictions daily and backfilling scores after each game. Across 24 NBA games: 0 exact-score hits, with the win/loss direction correct in 13 (54.2%). This is a no-spin retrospective with real numbers for anyone curious about what AI can actually do in sports prediction.

How the Numbers Are Calculated

When each prediction is published, I record: home team, away team, tip-off time, and the model's predicted score. After the game ends, I pull the actual score from api-football and auto-settle into three tiers:

• Exact score: predicted score = actual score (rare)
• Outcome correct: win/loss direction correct, score wrong (the main NBA tier)
• Miss: win/loss also wrong

Detailed methodology is in the "How We Predict" section of /predictions/.

The Truth About NBA: Exact Score Isn't a Reasonable Target

In soccer, low scores like 1-0 or 2-1 are everywhere, and the odds of AI guessing one isn't terribly low (9.9% sitewide). An NBA score combination like 110 to 105 has a space dozens of times larger than soccer—theoretically AI could hit one occasionally, but across 24 actual games, I got 0. The conclusion is simple: for NBA, just look at win/loss, don't look at the score.

That's why the hero block on /predictions/nba/ openly displays "Exact-score hit rate 0%, win/draw/loss hit rate 54.2%"—that's the ceiling on AI's ability to predict full scores.

Is 54.2% Win/Loss High or Low?

Long-run NBA public market (Vegas line) data shows blind betting on away wins around 42% and on home wins around 58%. The simple "home win = win" baseline = 58% accurate. My 54.2% is actually slightly below this brain-dead baseline (about 4 percentage points lower).

One interpretation is that the sample is too small (24 games, the first complete month). A more honest interpretation: the current prompt under-models NBA "home-court motivation / late-season state" and gets dragged down by baseline anchoring. Whichever interpretation you take, this AI prediction setup currently doesn't beat a strategy you could compute with one line of SQL—and that's something I won't wait to admit until "the numbers look better."

What's more, sportsbooks take ~5% vig, so to break even you need at least 52.4% accuracy (against -110 odds). The chance of AI beating sportsbooks purely on hit rate currently looks nonexistent. If anyone tells you they can win consistently at this, you shouldn't believe them.

My reason for doing this isn't to win money—it's to see where the limits are for AI in a "public data + structured analysis" workflow. Compared to buying courses, listening to experts, or chasing those "insider tips" on social, publicly admitting whether each prediction was right or wrong and exposing it for anyone to check is itself rarer than hit rate.

A Few Interesting Cases

24 games isn't many, but several typical AI failure patterns are already visible. Each is linked to the original; scores and analyses are all publicly verifiable.

1. 1-point squeaker, AI predicted reverse direction by +8 —— Knicks vs Hawks (4/21). AI gave Knicks a 109-101 home win by 8, but the actual result was 106-107, a 1-point Hawks road squeaker. 1-point games are normal in the NBA (most go to OT), and AI got neither the direction nor the spread—the most classic double-miss.

2. A 33-point swing massacre —— Trail Blazers vs Spurs (4/26). AI predicted a 12-point home win (118-106), but the home team actually lost by 21 (93-114). A 33-point error—AI completely missed the Spurs' form that night.

3. Spread nailed perfectly—score off but spread perfect —— Spurs vs Trail Blazers (4/20). AI predicted 118-105 (spread +13), actual 111-98 (spread +13). Score was off but the spread was nailed exactly—the cleanest "methodology worked" moment across 24 games.

4. Lakers blow out Suns by 28 —— Lakers vs Suns (4/11). AI gave Lakers a 7-point win (115-108), but Lakers actually won by 28 (101-73). Suns scored 73 for the entire game (league average is 113), suggesting AI didn't have the Suns' injury or rest situation that night—exactly the kind of private information a prompt can't access.

5. Both games AI picked road wins, the home teams blew them out —— Magic vs Pistons (4/25) + Rockets vs Lakers (4/27). AI picked road wins for both, and the home teams blew them out (home wins by 8 / 19). This is the most consistent systemic AI bias in the data: under-fitting on home-court motivation.

Every entry can be traced through /predictions/nba/ to the next one—I haven't hidden anything after winning a game; all settlement results are listed in the sitemap under the _match_settled tag.

What I've Learned

1. Score prediction isn't a product direction: Asking AI to bet on a six-figure combination like 110-105 is a meaningless precision illusion. It should pivot to win-rate distributions or over/under—questions with probabilistic edges.
2. Prompt engineering isn't monotonically increasing: A more complex prompt isn't necessarily more accurate. Early on I used minimalist prompts and the hit rate wasn't bad; after stuffing in a bunch of player props, the hit rate didn't visibly lift either. Now I run A/B tests to choose between prompts.
3. Publicly owning missed games matters more than pointing out hits: Weekly postmortems run automatically, writing the reasons hit rate dropped back into the prompt feedback. Without forcing this in public, model iteration becomes selective storytelling.

Next Steps

• 6 league hubs are live (NBA / EPL / Serie A / La Liga / UCL / CSL)—each filtered by real hit rate per league
• TG channel @joysport_predictions for real-time updates—feel free to follow along

If you're also experimenting with AI + public data, hit rate isn't what matters most; publicly putting your hit rate on the home page is.

Disclaimer: This article does not constitute betting advice in any form. Please comply with your local laws and regulations.