How we test it — and what the numbers actually mean.
This page explains how the engine builds a ticket set and how we tested it against random selection, with the current figures.
The engine uses 15 statistical models with equal weight. The historical replay below measures the combined engine, not the models one by one; nothing is treated as proven.
The replay scores each historical draw using only the draws before it. It is a replay of past data, not a record of forecasts made at the time.
To build a set, the engine draws tickets from one scored number pool, applies structural filters and a fixed popularity heuristic, and keeps the combination that best balances spread, ticket score and mutual overlap. It has no data about what other players actually play.
We don't predict winning numbers. We build reviewable sets.
The V4 Engine does not predict winning lottery numbers.
Its purpose is to build a full ticket set in a repeatable, inspectable way and to report the set's measurable properties next to those of a random set of the same size. This historical backtest detected no reliable predictive advantage over random selection.
The structural metric we measure is coverage: how widely a ticket set spreads its selections across the number pool, expressed as the average number of main numbers any two tickets in the set share. Fewer shared numbers means wider coverage. A random set of 5-of-50 tickets shares 0.50 numbers per pair on average.
Coverage is a property of the set, not a prediction. We report it next to the match rate and the winning-number rank so that each can be compared with its random expectation on the same draws.
Historical replay (walk-forward)
A replay can be fooled by hindsight if a model is scored on draws it has already seen. Ours is not: for every target draw, the engine sees only the draws before it.
The replay runs the engine on the draws preceding each target draw, then scores that draw. It is a replay of past data, not forecasts recorded at the time, and the model set was chosen in May 2026 using part of this same history, so the whole period is not untouched by model selection.
15 models. One ensemble score.
The V4 Engine combines 15 statistical models, each given equal weight, into a single ensemble score.
Each model looks at the draw history from a different angle: how often a number has appeared, how long since it last appeared, short-window versus long-run behaviour, which numbers tend to follow or accompany it, and how well it fits the typical shape of a draw. Three of the fifteen use no draw history at all; they are fixed shape preferences. The outputs are standardised and averaged. The replay measures the combined engine, not the models one by one; the combined engine showed no reliable advantage over random selection.
The numbers, simply stated.
The replay covers 440 Eurojackpot draws from 28 June 2022 to 11 September 2026, after 38 history-only warm-up draws, from a dataset of 478 draws. For each draw the engine used only the draws before it. Figures below are from the seeded run of 21 September 2026 (harness seed 20260915, engine seed 20260915, three evaluation runs per draw, balanced 12-ticket sets).
Corrected 21 September 2026: the 15 September benchmark supplied draw history oldest-first; the engine expects newest-first. The figures above come from the corrected run. This evaluation still does not demonstrate a predictive advantage.
This historical backtest detected no reliable predictive advantage over random selection. The drawn numbers sit at a mean rank of 25.19 in the engine's scoring against a theoretical expectation of 25.5, and engine tickets matched 0.51 main numbers per ticket against 0.50 for simulated random tickets. A mean of 0.51 correct numbers per ticket is a descriptive average over the replay, not a probability of winning anything.
Balanced 12-ticket engine sets are more concentrated than random sets: their tickets share 0.73 main numbers per pair against 0.50, because every line is drawn from the same scored pool. Set size and mode change this figure: a balanced 3-ticket set shared about 0.164 numbers per pair against 0.492 for random. These overlap figures describe ticket-set concentration; by themselves, they do not establish an improvement in winning chances or payouts.
The popularity score is an internal heuristic about common picking habits: numbers in the 1–31 birthday range, consecutive pairs, multiples of five, culturally popular numbers and evenly spaced patterns. It uses no information about what other players actually play, so no reduction in prize sharing is claimed. This evaluation does not establish that the engine improves your odds. It provides a repeatable, inspectable way to build a set and review it after the draw in the Virtual Vault.
Lottery draws remain fundamentally random events, and no system can guarantee a jackpot win.
The V4 Engine is designed to build structured, reviewable ticket sets — not to overcome the mathematical odds of the game itself.
Its goal is to make ticket-set construction explicit and testable, so that every claim on this page can be re-checked against the published protocol.
Your first full prediction set is free.
No card required.
Generate my AI set →