How We Built an AI That Optimizes Its Own Betting Strategy
1,290 experiments. 20,000+ historical bets. From -5% ROI to +20%.
12 min read
The Problem: +EV Bets That Weren't Profitable
Here's something most betting analytics platforms won't tell you: identifying positive expected value bets is not the same as making money.
Our EV pipeline was doing exactly what it was designed to do — calculating fair odds, removing the vig, and surfacing bets where the sportsbook's price exceeded our modeled probability. Textbook +EV methodology. The math was correct.
The results were not.
The Raw Numbers
• 20,000+ bets flagged as positive EV across multiple sports and seasons
• -5% ROI if you flat-bet every single one
• Meaning for every $100 wagered, you lost $5 on average
The EV model was finding edges that existed in theory but evaporated in practice. Some bets were priced as +EV because the odds were wildly mispriced on longshots that almost never hit. Others were in markets where the "sharp" consensus line was itself unreliable. The signal was real, but it was buried in noise. We needed a way to separate them — automatically.
The Karpathy Autoresearch Inspiration
The idea came from Andrej Karpathy (founding member of OpenAI, former head of AI at Tesla). He described a concept called autoresearch — a loop where an AI agent:
- 1Reads existing code and results
- 2Proposes a modification to a parameter or filter
- 3Runs the experiment against historical data
- 4Checks the target metric (in our case, ROI)
- 5Keeps the change if the metric improved, discards it if not
- 6Repeats
It's gradient descent for strategy. No human intuition, no "I think NFL spreads work better" — just relentless empirical testing. We adapted this concept specifically for bet selection parameter optimization and pointed it at our full historical dataset.
The system ran 1,290 experiments autonomously. Each one tested a hypothesis — "What if we cap odds at +200?" or "What if we exclude totals markets for NBA?" — and measured the impact on realized ROI.
The ~15 Parameters We Optimized
The pipeline didn't optimize one magic number. It explored a multi-dimensional parameter space, testing roughly 15 different levers that control which bets make it through our filter:
Odds & Probability
- Maximum odds cap (e.g., no bets above +300)
- Minimum odds floor
- Implied probability bands
- EV percentage thresholds (minimum and maximum)
Market & Sport
- Market type restrictions (moneyline, spread, total)
- Sport-specific inclusion/exclusion rules
- Side restrictions (home/away, favorite/underdog)
- Sport + market combinations
Bookmaker & Source
- Bookmaker preference ordering
- Book-specific EV threshold adjustments
- Line source weighting
Timing & Context
- Time-before-game windows
- Day-of-week filters
- Season phase adjustments
Every parameter was tested independently and in combination. The pipeline didn't just find the best single filter — it found the best combination of filters that worked together.
Key Discoveries
1. The Inverted EV Curve
This was the most counterintuitive finding: higher EV percentage does not mean higher ROI. In fact, the relationship is inverted at the extremes.
• Bets with 3-6% EV: Modest but consistent positive ROI
• Bets with 7-12% EV: ROI starts declining
• Bets with 15%+ EV: Deeply negative ROI
Why? Because extremely high EV percentages almost always correspond to longshot odds. A bet at +500 showing 20% EV sounds incredible on paper. In practice, the "true probability" estimate is unreliable at those odds levels. The model thinks there's a 22% chance; reality says it's closer to 14%. The vig removal math amplifies errors in thin markets. The lesson: moderate edges on likely outcomes beat large edges on unlikely ones.
2. Longshots Are Toxic
The data was unambiguous. Bets above approximately +250 odds destroyed returns regardless of the EV percentage. This aligns with well-documented academic research on the "favorite-longshot bias" — bettors systematically overvalue longshots, and sportsbooks know it.
The optimization pipeline discovered this independently. No one told it about the favorite-longshot bias. It just observed that capping maximum odds dramatically improved realized ROI, and it kept tightening the cap until it found the sweet spot.
3. Profitable Niches Exist — But They're Sport-Specific
Not all sports and market types are created equal. The pipeline found dramatically different profitability profiles depending on the sport-market combination:
NFL Moneylines
+38% ROI
NHL H2H
+10% ROI
NBA Spreads
+0.6% ROI
The takeaway: a single "one size fits all" EV threshold is leaving money on the table. NFL moneylines can tolerate different parameters than NBA spreads. The best strategy isn't a universal filter — it's a collection of sport-specific rules.
4. The "Ultimate Filter" — Sport-Specific Rules
The final output of the optimization pipeline wasn't a single threshold. It was a set of sport-specific filter rules, each tuned to the characteristics of that market. Different EV minimums per sport. Different odds caps. Different market type inclusions.
Combined Result
• Before optimization: -5% ROI on 20,000+ unfiltered bets
• After optimization: +20% ROI on the filtered, curated bet set
That's not a marginal improvement. It's a 25-percentage-point swing from applying data-driven filters to the same underlying EV pipeline. The math didn't change. The curation did.
Walk-Forward Validation: Proving It's Not Overfit
Any optimization pipeline can achieve great numbers on the data it trained on. That's trivial and meaningless. The only question that matters is: does it work on data it has never seen?
We used walk-forward validation — the standard method for testing time-series strategies without look-ahead bias:
- Train on historical months
The pipeline optimizes filter parameters using past bet outcomes
- Test on the next month (unseen)
Apply those filters to the following month's bets — data the optimizer never touched
- Slide the window forward and repeat
Move to the next period, retrain, retest — no peeking
The Result
The optimized filter set was profitable in every single out-of-sample month. Not on average. Not in aggregate. Every month. That's the strongest signal you can get that the edge is real and not an artifact of curve-fitting.
How Overfitting Almost Fooled Us
Midway through the optimization process, the pipeline produced a filter set that looked incredible: +34% ROI. We almost shipped it.
Then we looked at the sample size: 152 bets. All football. All from a narrow window where a specific bookmaker had consistently mispriced NFL and college football lines.
The 34% ROI Trap
• In-sample ROI: +34% (looks amazing)
• Sample size: 152 bets (far too small)
• Sports: Football only (too narrow)
• Out-of-sample: Inconsistent — profitable some months, deeply negative others
This is a textbook overfitting scenario. The optimizer found a pattern that existed in one specific slice of historical data but didn't generalize. With only 152 bets, the confidence interval around that 34% ROI was enormous — it could easily have been luck.
We built minimum sample size requirements into the pipeline after this. Any filter set that couldn't produce at least several hundred qualifying bets per evaluation period was automatically rejected. The 20% ROI filter we shipped is based on a much larger sample, works across multiple sports, and — critically — survived walk-forward validation.
What This Means for +EV Bets Users
Every bet you see on +EV Bets passes through these optimized, sport-specific filters. The raw EV pipeline finds thousands of potential edges every day. The autoresearch-optimized filter layer ensures you only see the ones that have historically translated into real profit.
1,290 Experiments
Every filter was earned through empirical testing, not guesswork
Sport-Specific Rules
Not a one-size-fits-all threshold — optimized per sport and market type
Walk-Forward Validated
Profitable in every out-of-sample month — not curve-fit to history
See the Optimized Picks
Start your 7-day free trial. Every pick is filtered through 1,290 experiments.
Frequently Asked Questions
What is the autoresearch optimization pipeline?
It's an automated system inspired by Andrej Karpathy's autoresearch concept. The pipeline modifies bet selection parameters, runs backtests against historical outcomes, evaluates the results, and keeps or discards each change — repeating this loop hundreds of times to converge on optimal filters.
How many experiments did the system run?
The pipeline ran 1,290 experiments across 20,000+ historical bet outcomes, testing combinations of odds caps, market restrictions, EV thresholds, side restrictions, bookmaker preferences, and implied probability bands.
What was the ROI improvement?
The unfiltered bet pool had a -5% ROI across all 20,000+ bets. After optimization, the curated filter set achieved +20% ROI on in-sample data and remained profitable across every out-of-sample month in walk-forward validation.
How do you prevent overfitting?
We use walk-forward validation — the model is trained on historical data, then tested on future months it has never seen. A filter set must be profitable in every out-of-sample period, not just on average. We also require minimum sample sizes to avoid strategies that look great on a handful of bets.
Why does higher EV not mean higher ROI?
Counterintuitively, bets with very high EV percentages (15%+) tend to be longshot bets where the sportsbook has mispriced odds dramatically. These bets hit rarely, and the variance is extreme. The profitable sweet spot is moderate EV on shorter-odds bets where the edge is real and the win rate is sustainable.
Past performance does not guarantee future results. Always bet responsibly.