
How to Evaluate AI Picks Before You Bet on Them

The best way to evaluate AI picks is to judge the record, not the pitch. A credible AI pick service shows every pick with the odds it was graded at, grades it publicly, keeps a permanent archive, and reports results in units across a large sample. Win rate alone tells you almost nothing. Manny’s Variety is built around that standard: a public, permanent record of every pick, graded in the open, with results reported in units.
This guide gives you the exact checks, the math behind each one and worked examples you can repeat with any service.
TL;DR
- Win rate without odds is meaningless. A service can win 81% of its picks and still lose money if it only backs heavy favorites.
- Judge results in units and return on investment (ROI), with the price each pick was graded at.
- Sample size matters: 50 picks prove almost nothing, and 1,000 or more start to say something.
- Look for a permanent, public archive that includes losses, not screenshots of winners.
- Check calibration and closing line value (CLV) when the service publishes enough data to do it.
Table of Contents
- Why a headline win rate is not enough
- The five numbers to check on any AI picks record
- How to evaluate AI picks step by step
- Calibration: does a 64% pick win about 64% of the time?
- Closing line value: beating the market price
- Warning signs in an AI picks record
- Comparing picks across sports, bet types and moving odds
- Our view on judging AI picks
- What Manny’s Variety gives you to work with
- FAQ
- Sources
Why a headline win rate is not enough
Most AI pick pages lead with one number: “63% winners” or “80% accuracy.” That number is only half a result. A bet pays according to its odds, so the same win rate can produce a profit or a loss.
Here is the standard example. Say a tipster backs short-priced favorites and wins 81% of the time at an average decimal price of 1.18. Each $1 staked returns 0.81 × 1.18 = $0.956, which is a loss of about 4.4% on every dollar. B337 walks through the same arithmetic and recommends judging a tip service on a dated record of every pick rather than on accuracy figures or screenshots.
The reverse also holds. A service that wins only 45% of its bets can be profitable if it backs underdogs at long enough odds.

The break-even point for standard odds
Most US point spreads and totals are priced at -110, meaning you risk $110 to win $100. The break-even win rate at -110 is 110 ÷ 210, or 52.4%. Any win rate you see on a spread or totals record should be compared with that number, not with 50%.
For moneylines and props the break-even changes with every price. That is why the record must show the odds, not only the result.
The five numbers to check on any AI picks record
1. Win rate, with the odds beside it
Win rate is useful once you know the typical price. At -110, a 63.5% win rate over 100 flat one-unit bets works out like this:
- 63.5 wins × 0.909 units = 57.7 units won
- 36.5 losses × 1 unit = 36.5 units lost
- Net: +21.2 units, a 21.2% return on units staked
That is the kind of arithmetic you should be able to run on any record. If the service does not publish prices, you cannot run it.
2. Units won or lost
A unit is a fixed stake size, such as 1% of a bankroll. Reporting in units lets you compare services without knowing anyone’s bankroll. It also puts wins and losses on the same scale: a loss at -110 costs one unit, and a win pays 0.909 units.
Manny’s Variety reports its record this way. Its picks have a 63.5% win rate across more than 1,600 tracked picks, with a net return of 443.9 units, and each pick sits in a permanent archive.
3. Return on investment
ROI is net profit divided by total amount staked. For decimal odds with fixed-unit stakes, iSportsAPI gives the formula as the sum of each bet’s winnings minus its stake, divided by the sum of stakes. In plain terms: add up profit, divide by money risked.
A simple spreadsheet is enough to track it. Our guide to measuring sports betting ROI with a spreadsheet shows the columns to keep and how to read the result once you pass a few hundred bets.
4. Sample size
Small samples flatter everyone. Take a service that wins 60% over 50 picks. The 95% confidence range around that figure (a standard statistical estimate, using the usual normal approximation) runs from roughly 46% to 74%. The lower end is below the 52.4% break-even. The same 60% over 1,000 picks narrows to about 57% to 63%, which sits clearly above break-even.

A practical rule: treat anything under a few hundred picks as an early sign, and wait for 1,000 or more before drawing a firm conclusion. A record of more than 1,600 picks is in the range where the numbers carry weight.
5. How the record is kept
The quality of the record matters as much as the numbers inside it. The questions are simple:
- Is every pick published before the game starts, with a timestamp?
- Are losses and pushes in the same list as wins?
- Can old picks be edited or deleted?
- Is grading done in public, against stated odds?
Manny’s Variety keeps a permanent archive of every pick and grades results publicly, so the record you read today is the one that will be there next season. Our article on what a credible sports picks archive shows covers each element in detail.
How to evaluate AI picks step by step
Use this routine on any service, including ours. It takes about 20 minutes with the archive open.
- Find the full list of picks. You want every pick, not a highlights page. A good archive lets you scroll back months or seasons.
- Check that picks were posted before the event. A pick with no timestamp can be written after the result.
- Note the odds on each pick. Without prices, you cannot compute units or ROI. This is where most weak records break down.
- Count the sample. Total the graded picks. Under a few hundred, hold your conclusions loosely.
- Compute win rate, units and ROI yourself. Do it for at least one sport and one bet type. If the service’s own totals match your sums, the record is internally consistent.
- Split by sport and bet type. An AI engine can be strong in the NBA and average in props. Splitting shows where the edge sits. Manny’s Variety runs sport-specific engines, so the archive lets you read each sport on its own.
- Look at recent results against long-run results. Variance produces hot and cold stretches. Compare the last 100 picks with the full record before judging a streak.
- Decide your stake size in units. Flat staking, such as 1 unit per pick, keeps results comparable to the published record.
Calibration: does a 64% pick win about 64% of the time?
Some AI services publish a probability or confidence for each pick. That gives you a second test beyond win rate: calibration. Among all picks labeled around 64%, roughly 64% should win.
The standard measure is the Brier score, the average of (forecast − result)², where a win counts as 1 and a loss as 0. B337 gives the example: a 64% pick that wins scores 0.1296, and one that loses scores 0.4096. Lower is better. The score is only meaningful against a baseline, so compare it with the margin-free chances implied by the market on the same games.
You do not need statistical software for a rough check. Group picks by stated confidence (for example 55-60%, 60-65%, 65-70%), then compare each group’s win rate with its label. A model that says 70% and wins 55% is overconfident. A model whose groups track their labels is giving you probabilities you can size stakes around.
Closing line value: beating the market price
The closing line is the final price before the game starts, and it reflects the most information the market will have. If you consistently take a better price than the close, that is an early sign of skill that does not depend on short-term luck.
iSportsAPI defines CLV as (odds taken ÷ closing odds − 1) × 100. A simple example in decimal odds: you take 2.10 and the line closes at 2.00. That is (2.10 ÷ 2.00 − 1) × 100 = +5% CLV on that bet.
Treat CLV as a signal, not a guarantee. It is an early indicator that a model may be finding market inefficiencies, and it does not by itself prove long-term profit. Units and ROI over a large public sample remain the final measure. Use CLV when you can track it on your own bets: note the price you got, then the closing price, and average the difference over time.
Warning signs in an AI picks record
Some signs a record is incomplete or shaped to look better than it is:
- No odds on picks. Results without prices cannot be turned into units.
- Screenshots instead of an archive. A screenshot shows one result, chosen after the fact.
- Picks that appear only after they win. Look for timestamps before game time.
- A record that starts after a hot streak. Ask when tracking began and whether earlier picks exist.
- Backtests presented as live results. A backtest runs a model over past games. It can suffer from look-ahead bias (using information that was not available at the time) and from tuning the model until it fits history. Live, graded picks are the stronger evidence.
- Only the best sport shown. If a service covers six sports, ask for all six.
- Accuracy claims with no sample size. “80% accurate” over 20 picks is noise.
When you see several of these together, the record is not one you can verify. A public, permanent, fully graded archive removes most of them at once.
Comparing picks across sports, bet types and moving odds
Past results do not guarantee future profit, so use them to compare, not to predict. A few habits keep the comparison honest.
Compare like with like. A moneyline record, a spread record and a player props record have different break-even points and different variance. Read each on its own. Parlays compound variance, so a parlay record needs more picks before it says much.
Track the price you actually got. Lines move. A pick graded at -110 may be -125 by the time you place it. Your results follow your price, so shop for the best number and log it.
Read by sport. A sport-specific engine can have different strengths in different leagues. Our pages on NBA AI predictions and NHL AI predictions show how picks are presented for each sport.
Check consensus against the record. A pick that many sources agree on is not automatically a good one. The six steps in our guide to verifying consensus and AI picks with records apply the same test to NFL picks.
Keep the sample growing. Re-run your numbers every 100 picks or so. If the record holds across seasons and sports, the evidence compounds.
Our view on judging AI picks
We built Manny’s Variety on one rule: a pick is only worth something if anyone can check it afterward.
That is why every pick goes into a permanent archive and is graded in public. Our sport-specific engines use real-time analytics and thousands of simulations for each sport, and the output is held to the same standard as any bettor’s own picks: units, odds and a growing sample.
The numbers we publish are 63.5% winners across more than 1,600 tracked picks, with a net return of 443.9 units. We want you to run the math on them the way this guide describes, in your own spreadsheet, and compare results by sport.
A record you can verify lets you size your stakes with confidence and judge the service on evidence. That is what an amateur bettor needs most.
Manuel
What Manny’s Variety gives you to work with
If you want a service built around the checks above, Manny’s Variety offers three plans on its pricing page:
- Free: $0, with limited features.
- Core: $39.99 per month.
- Elite: $79.99 per month.
The platform covers MLB, NBA, NFL, NHL, PGA, MMA, soccer and international tournaments, with daily reports, market analysis, player props, parlays and fantasy football tools. Short-term day and week passes are also available.

Start with the free plan and the public archive, then compare your own spreadsheet with the published record. If you are still choosing a platform, our verifiable checklist for choosing a sports betting platform lists what to look for before you pay.
FAQ
What is a good win rate for AI sports picks?
It depends on the odds. At standard -110 pricing the break-even rate is about 52.4%, so anything above that over a large sample is profitable before fees. Judge the win rate together with units won and the number of picks. A high win rate on short-priced favorites can still lose money.
How many picks do I need before I can trust a record?
A few hundred picks give an early read, and 1,000 or more give a much firmer one. With only 50 picks, a 60% win rate could easily be luck, because its 95% range runs from about 46% to 74%. More picks narrow that range.
Should I trust an AI pick service that shows screenshots of winning bets?
Screenshots show single results chosen after the fact. A permanent archive that lists every pick, wins and losses, with timestamps and odds, is the stronger form of evidence. That is the kind of record Manny’s Variety keeps.
What is closing line value and do I need it?
CLV compares the price you took with the final price before the game. Beating the close consistently is an early sign of skill. You do not need it to start, since units and ROI over a large sample answer the main question, but it is a useful extra to log on your own bets.
Can I evaluate AI picks without advanced math?
Yes. Log each pick’s odds and result in a spreadsheet, total the units won and divide by units staked. That gives you ROI, and the steps above cover the rest. Our spreadsheet guide to sports betting ROI shows the layout.
Sources
- AI Sports Betting: How to Test AI Predictions and Tips (B337)
- Sports Prediction Models: ROI, CLV & Rolling Window Backtesting (iSportsAPI)