Do Lottomatikai data guarantee future wins?

No. Lottomatikai provides historical statistical analysis of past draws. Past performance does not guarantee future results. Lottery games have intrinsic negative expected value.

How does Lottomatikai's AI work?

We use an ensemble of machine learning models (Random Forest, Gradient Boosting, Bayesian priors) trained on the complete history of Italian draws. Models identify statistical patterns, not deterministic predictions.

Can I recover losses with Lottomatikai?

No. The expected value of lottery games is negative: on average you lose. Only play amounts you can afford to lose. If gambling becomes a problem, contact responsible gambling operators.

What does AI accuracy 12% vs random 5.5% mean?

It means that, in the last 90 days, AI predictions achieved an average hit rate of 12% compared to the 5.5% expected from random selection. The data is statistically significant (p<0.05) but remains historical analysis, not future prediction.

Who produced the predictions shown in the trophies?

They were generated by the Lottomatikai AI system before the draw, and their outcome is verified after the draw. They are not plays made by customers, nor results reported by third parties: they are verified outcomes of system predictions. Trophies remain completely anonymous: no name, email or identifier is shown, only tier labels (AI Elite System, AI Pro System, Virtual Journal).

Is Lottomatikai authorized by ADM?

Lottomatikai is a statistical analysis tool. It is not a gaming operator and does not offer direct plays. Plays must be made at ADM authorized operators through physical retailers or online concessionaires.

How reliable are the ensemble model weights?

Weights are periodically calibrated via meta-learner on rolling windows. On 10eLotto we use a conformal set: the membership threshold comes from the α-quantile on the calibration sample, and a number is marked “in set” only if it exceeds it. On the other games the conformal set is not computed, and indeed it is not shown. Empirical coverage has not yet been verified out of sample: once it is, the figure will be published here.

Can I trust the per-tier simulations on /pricing?

Simulations are based on real predictions aggregated per tier from the last 90 days. We show only tiers with at least 30 predictions to ensure statistical significance. They are historical analysis, not promise of return.

Correction to historical data

The figures on this page have been recalculated. Three corrections, all downward: plays still awaiting their draw are no longer counted as losses; rows written AFTER the draw that evaluates them are excluded (they were historical backfills, not predictions); and the comparison against chance now uses the probability of the play actually proposed, not that of a harder combination. The success criterion is single and stated: the play collected a paying category. Winning does not mean being in profit — EV remains negative by definition.

Updated on August 4, 2026

Lottomatikai AI Track Record

Q: Who produced the predictions shown in the trophies?

They were generated by the Lottomatikai AI system before the draw, and their outcome is verified after the draw. They are not plays made by customers, nor results reported by third parties: they are verified outcomes of system predictions. Trophies remain completely anonymous: no name, email or identifier is shown, only tier labels (AI Elite System, AI Pro System, Virtual Journal).

Q: Is Lottomatikai authorized by ADM?

Lottomatikai is a statistical analysis tool. It is not a gaming operator and does not offer direct plays. Plays must be made at ADM authorized operators through physical retailers or online concessionaires.

Q: How reliable are the ensemble model weights?

Weights are periodically calibrated via meta-learner on rolling windows. On 10eLotto we use a conformal set: the membership threshold comes from the α-quantile on the calibration sample, and a number is marked “in set” only if it exceeds it. On the other games the conformal set is not computed, and indeed it is not shown. Empirical coverage has not yet been verified out of sample: once it is, the figure will be published here.

Q: Can I trust the per-tier simulations on /pricing?

Simulations are based on real predictions aggregated per tier from the last 90 days. We show only tiers with at least 30 predictions to ensure statistical significance. They are historical analysis, not promise of return.

Historical statistical analysis of AI predictions — past performance does not guarantee future results.

Stated criterion and sample · verifiable computation

Campione insufficiente. I dati si popolano automaticamente con l’accumulo di estrazioni.

View all verified plays →

Our position

On these games draws are independent: no model can pick numbers better than chance, and anyone claiming otherwise is not measuring it. What we do is state every figure with its margin of error and let you check it. Where an edge does exist is elsewhere: on SuperEnalotto, which splits the pot among winners, choosing rarely played combinations does not raise your chance of winning — it raises how much you collect when it happens, because you share with fewer people.

How to read this

LIVE signals: our real signals, issued before each draw and evaluated afterwards.
Historical simulation: the same strategies replayed on past draws (backtest), for a large, honest sample.
What it is for: we do not promise wins: expected value stays negative. We help you play adequately — size your stake, contain losses, avoid burning your budget, and cash out when you win.

HISTORICAL SIMULATION

Historical Simulation

Simulated signals

0.0%

Hit rate

Full data (incl. P&L simulation)

+0.00 €

Cumulative P&L

+0.0%

ROI

Dig into the track record

The same evidence, broken down by game, by strategy and over time.

Which games perform bestROI, lift and hit rate compared across games.Open Which strategies work bestRisk-return profile and calibration per strategy.Open Does the model self-improve?Calibration evolution over time and room for improvement.Open All verified playsHistorical archive of hits, with filters by game and period.Open

View all verified plays →

The engine

Does the engine, every day, select better than chance?

Frequenza osservata contro probabilità teorica, per gioco

Scarto fra la frequenza osservata dell’evento «almeno un numero predetto è stato estratto» e la sua probabilità esatta, che su questi giochi non dipende da quali numeri vengono scelti. Linea rossa = riferimento zero. Non è un vantaggio sul banco e non è la vincita — quella è misurata dal margine di selezione qui sotto: uno scarto positivo su poche estrazioni è quasi sempre varianza, e infatti sopra o sotto viene dichiarato solo quando supera l’errore statistico.

Caricamento lift...

Calculation windows are rolling over 90 days: two consecutive points share 89/90 of the sample, so the 82 points above are not 82 independent observations — they are the same series seen one day apart. The largest deviation measured this way (SuperEnalotto, 90-day window) is −1.4σ, p ≈ 0.17: not statistically significant.

What you actually received

What would someone who followed the real-money advice have earned?

The plays recommended with a real stake, one per draw: a much smaller sample than the engine's above, but the only one that counts a real cost and a real payout.

See the trophies and the public bankroll →

What we genuinely guarantee

What is true by construction, before any draw takes place?

No one else shows you this table.

No sample: it's a theorem, not a measurement.

measurement (with interval)

Splitting the prize

It doesn’t increase your odds of winning — it increases how much you take home if you do, because it avoids the numbers everyone plays: in out-of-sample testing it passed 10 out of 10 checks by avoiding the most-played numbers (11 is played 1.51x the average, 61 only 0.74x).
It reduces the risk of your win being split: Italy’s MD rules cap identical tickets at 200 per draw — beyond that, the prize is shared among winners of the same combination — and a less-played combination lowers the odds of crossing that threshold.
On Lotto the payout is fixed, so it doesn’t increase how much you take home no matter which numbers you choose: this lever doesn’t exist here.
On 10eLotto the payout is fixed, so it doesn’t increase how much you take home no matter which numbers you choose: this lever doesn’t exist here, for the same reason as Lotto.

theorem

The combinatorial guarantee

With these 12 numbers across 8 lines, if at least 4 of your chosen numbers are drawn, you certainly get at least a pair. It’s a theorem verified by exhaustive enumeration, not a promise: before the draw we don’t estimate the guarantee — we prove it, or we refuse to state it.

It doesn’t guarantee your numbers get drawn, nor a net win: the cost of the system stays at a negative EV by definition.

cost

Cost and survival

At a constant negative EV, the shape of your play changes variance, weekly spend, and time to bankroll exhaustion: choose how many numbers, how many wheels or lines, and see the numbers recalculated live — never a second prize table.

theorem

Calibration

Across all four games the probability we serve is known by construction — not estimated, not calibrated after the fact: 5/90 on Lotto, 1/90 on the Gold Number, 1 in C(90,6) on SuperEnalotto sixes, the same constant odds on 10eLotto and MillionDAY. When a probability doesn’t vary from sample to sample there is nothing left to compare it against: it’s declared non-calibratable, and that’s a permanent condition, not a measurement pending. That the number we serve is the true one — not a normalized share that sums to something other than one — is already the guarantee.

Past performance does not guarantee future results. · No guarantee of winning. · The expected value is structurally negative. · 18+ · adm.gov.it

The engine

Does the engine, every day, select better than chance?

Frequenza osservata contro probabilità teorica, per gioco

Caricamento lift...

What we genuinely guarantee

What is true by construction, before any draw takes place?

No one else shows you this table.

No sample: it's a theorem, not a measurement.

measurement (with interval)

Splitting the prize

It doesn’t increase your odds of winning — it increases how much you take home if you do, because it avoids the numbers everyone plays: in out-of-sample testing it passed 10 out of 10 checks by avoiding the most-played numbers (11 is played 1.51x the average, 61 only 0.74x).
It reduces the risk of your win being split: Italy’s MD rules cap identical tickets at 200 per draw — beyond that, the prize is shared among winners of the same combination — and a less-played combination lowers the odds of crossing that threshold.
On Lotto the payout is fixed, so it doesn’t increase how much you take home no matter which numbers you choose: this lever doesn’t exist here.
On 10eLotto the payout is fixed, so it doesn’t increase how much you take home no matter which numbers you choose: this lever doesn’t exist here, for the same reason as Lotto.

theorem

The combinatorial guarantee

It doesn’t guarantee your numbers get drawn, nor a net win: the cost of the system stays at a negative EV by definition.

cost

Cost and survival

theorem