Build your first betting model with AI. Then find out if it actually works.

I'm not a programmer. Back in May I started building a betting model with AI for two of my buddies, and the first thing I learned is the model isn't the hard part. Knowing if it actually works is.

So here's how to build your own in about an hour, and then find out for real. Not by wins. By the closing line.

Works for NHL picks and NBA moneylines and totals.

Step 1: Pick one belief

Every model starts with one sentence: the market gets this wrong, and here's why. Write yours down before you look at any data. Trust me on this one.

Keep it to one sentence. If it needs a paragraph, it's really three ideas.

Make it about the price, not the team. "The Avs are good" isn't a belief. Everybody knows the Avs are good, and the price already says so. "Home underdogs are priced too long" is a belief.

And you have to be able to check it before the game, from the schedule and the odds.

Here are three our own lab is racing right now:

• Early-season NBA totals are set too high.
• An NBA price that moved overnight isn't done moving by morning.
• DraftKings underprices NHL home underdogs.

Some of ours are ahead of the close. Some are getting smoked. You don't know until you race them. See them live.

Step 2: Turn it into a rule with AI

A rule is your belief written so tight that two different people would pick the exact same games. This is where AI is actually great. Paste this into ChatGPT, Claude, Grok, whatever you use, with your belief on the first line:

My belief: [your one sentence].

Turn it into a betting rule I can follow every day with no judgment calls.
- Sport: NHL (moneyline, totals or puck line) or NBA (moneyline or totals).
- Say exactly which games qualify, using only things I can see before the game: the schedule, the teams, and the DraftKings odds.
- Say which side to bet.
- Say the latest time I can post the pick.
- Give me a short checklist I can run each morning in under 10 minutes.
Do not add extra conditions to make it look better. Keep it simple.

Then ask it one more thing: "What would make this rule wrong?" Save that answer. That's what your grade is going to test.

Once it's written, freeze it. Don't touch it after a bad night. I know it's tempting. But if you keep changing the rule, nobody can grade it, because nobody knows which version lost.

Step 3: Post your picks on X

Post from your own account, before the game starts. Our bot reads your post like plain text, so the format matters more than you'd think:

1. One pick per line. Nothing else on the line.
2. The league on its own line, above its picks. NHL or NBA.
3. Team, then market, then price. The price is optional. We grade the DraftKings price near when you posted.
4. Bullets and numbers are fine. "Lock of the night" is not.
5. Don't edit it after. The first read is final.

A post that reads clean:

NHL
MTL ML -195
SJS/NSH over 6.5 -110
NBA
BOS ML -120

We don't grade parlays, props, alt lines, or NBA spreads. Full posting rules here.

Step 4: Get graded

1. Fill out the Get graded form. Pick "Builder", your sport and your markets. Two minutes, tops.
2. We email you a short code. Post it from your X account so we know it's really you, then reply with the link.
3. Your page goes live, and grading starts with your next pick.

It's free, and it stays that way. Nobody can pay us to get graded, change a grade, or get taken off. 21+.

Step 5: Read your grade honestly

Your page shows two things, and I'd watch the second one.

Your record is your wins and losses. Over 50 or 100 picks, honestly, a lot of that is luck.

Your average against the close is whether you got a better price than where DraftKings closed. That one is way harder to fake, and it shows up long before your record settles down.

You get a spot on the board after 25 graded picks. Until then your page says "too early to call", and it means it.

Me? I'll tail one of our lab rules sometimes, but only the ones with a lot of picks behind them. Never one that still says too early to call. Treat your own model the same way.

And if you're behind the close after 100 picks, that's an answer too. Write down what you learned, freeze a new belief, and race that one.

The whole test, start to finish, is on one page: How to test a betting model.

Mistakes I'd skip (I've made most of them)

• Asking AI to "find an edge." It'll hand you something that fits old games perfectly and means nothing going forward. Start with your belief, not its search.
• Stacking conditions. Every "and only when..." makes the past look better and the future worse.
• Tweaking after losses. Freeze it. A new idea is a new rule, raced from zero.
• Judging by your record. 12-5 can be pure luck. The close is harder to fool.
• Posting late. If the line already moved, you're graded on the price you could actually get.

Questions

Get graded