TRANSPARENCY · RESULTS · RISK CONTROL
July audit: what failed and what changes in Deep Tennis AI
Publishing only the final result would be the easy part. Here I separate the official figures from a retrospective simulation, explain which segments did the damage, and put in writing the rules that start being validated going forward.
Between July 1 and 22, Deep Tennis AI went through two realities at once: the Premium Picks stayed positive, but several segments of the overall set built up losses that were too large. Instead of hiding them, I used that data to reduce exposure and make the system more selective.
What exactly I measured
The query uses the production database closed as of July 22, 2026, Madrid time. It includes picks from the 10.8 system in BO3 and BO5. Only matches with a decided result enter the main calculation; pending and void ones are kept separate.
For Premium I respected the website's public cut and the official execution applicable on each date, including moneyline protection when it applied. ROI is calculated over all units exposed. Drawdown measures the maximum drop in the curve from a previous peak.
Premium: official result versus the simulation
277u exposed. A theoretical 100u bankroll would have ended at 105.69u.
167u exposed. The same bankroll would have ended at 125.17u.
The simulated difference is +19,48u, but it doesn't come from “finding” winners after the match. It comes from applying the same market and risk rules to the entire sample: removing two specific segments and reducing the maximum stakes.
The two blocks that hurt Premium the most
- −1.5 sets between 1.90 and 2.09 odds: 11 picks, 3W–8L, −13.70u and −18.77% ROI.
- Game handicap on hard court: 4 picks, 1W–3L, −11.04u and −58.11% ROI.
The stake cap would also have reduced the profit from some strong winners. I won't hide it: the 4u limit isn't designed to maximize this backtest, but to prevent a maximum-conviction read from concentrating too much risk.
The full value picks book
The improvement is even more visible when looking at all the value picks that were publishable, not just Premium.
The simulation improves the result by 64.72 units. H2H outside clay contributed 30 picks, 11W–19L and −19.58u. The −1.5 in the 1.90–2.09 range contributed 23 picks, 6W–17L and −34.15u. Total games doesn't change this comparison because it was already outside July's public set.
In BO3 alone, the rules would have taken the result from −30.99u to +32.87u. The overall figure ends at +7.53u because BO5 keeps almost its entire universe and remains the block still pending a specific audit.
Deep Signals: a small but clean adjustment
Tapado IA used a confidence band from 60 to 64. The review showed that restricting it to exactly 60 confidence would have eliminated one loser:
- Old rule: 23W–18L, +2.73u, ROI +6.66% and DD −5.36u.
- New simulated rule: 23W–17L, +3.73u, ROI +9.33% and the same DD.
Winner Elite doesn't change: 16W–1L, +3.83u, ROI +22.53% and DD −1u. There was one pending Tapado IA, excluded from the closed results.
What changes from July 23
- BO3 H2H value picks can only be published on clay.
- BO3 −1.5 sets is blocked between 1.90 and 2.09 odds.
- Total match games stops being published as a BO3 value pick.
- The main stake has a global maximum of 4u; the game handicap keeps a stricter maximum of 2u.
- The game handicap on hard court cannot be Premium.
- Tapado IA requires confidence exactly 60 and odds between 1.50 and 2.50.
- CLV only compares the same market, selection and signed line; a −1.5 can never be crossed with a +1.5.
Premium protection keeps working as a separate entry when the pick meets its conditions. Winner Elite is not modified. Historical losses remain exactly where they were.
Why +25.17u isn't a promise
There's an obvious limitation: these filters were chosen after studying July and are tested on that same sample. That introduces selection bias. The numbers allow for a reasonable risk-control decision, but they don't prove future performance will be the same.
A backtest is useful for deciding what deserves a trial. Validation begins when the rule is frozen and matches we don't yet know about arrive.
That's why the changes are applied going forward and with review thresholds. I won't lift a veto after three wins or add another after three losses. A new, substantial sample is needed to check win rate, ROI, drawdown and stability across surfaces.
ONGOING TRANSPARENCY
Check the official results.
The backtest stays documented here; the public track record will keep showing only what actually happened.
Informational content for adults over 18. Betting involves risk of loss. Past results and retrospective simulations do not guarantee future results.