What Aleavici is, and what it is not

What we do

Aleavici builds forecasting models across football, financial markets, geopolitics and fiscal policy. Each runs the same way: public data is ingested nightly, features are rebuilt, models are retrained on a strictly chronological split, and the day's projections are published automatically. Nothing is hand-adjusted after the fact.

The football model is the most developed. It is trained on more than 275,000 played fixtures and produces, for every upcoming match, a probability for each of the three outcomes, expected goals for both sides, and an exact-score distribution drawn by Monte Carlo simulation.

How we know it works

Models are split by date, never at random: trained on fixtures up to 2023, tuned on 2024, and measured on 2025 — a season the model has never been shown. Measuring on data a model was trained on inflates every number and is the most common way a forecasting claim becomes meaningless.

On that held-out season, across 13,998 fixtures, the football model called 51.4% of three-way outcomes correctly against 33% for a coin. More useful than the headline accuracy is whether its stated confidence is honest:

• When it says 33–45%, the outcome happens 40.0% of the time (n = 5,074)
• When it says 45–55%, it happens 48.8% of the time (n = 4,144)
• When it says 55–70%, it happens 59.2% of the time (n = 3,058)
• When it says over 70%, it happens 77.1% of the time (n = 1,722)

A model whose stated probabilities match observed frequencies is calibrated. That is the property worth having, and it is the one we publish.

What these numbers cannot do

They cannot tell you what will happen. A 77% call is wrong roughly once in four, and that is the model working correctly rather than failing.

We do not publish betting advice, staking suggestions or claims of an edge over bookmakers. Where our probabilities differ from market prices we say so plainly and call it a disagreement, because that is what it is: two methods reaching different conclusions, with only the result able to settle which read it better. Anything stronger would be a claim about expected profit, and we have not established one.

Mistakes

Every prediction is recorded before kick-off and scored against the result once it is known. Where the model is wrong, it stays on the record. A forecasting site that only shows its successes is not showing you a track record, it is showing you a selection.

Publisher

Aleavici is published by Cosmopti, established in France.
Contact: contact@aleavici.com

Hosting: OVH SAS, 2 rue Kellermann, 59100 Roubaix, France — servers located in Montréal, Canada.