About Cricombo

What are the methods?

The whole point of the project was that I didn't want to lean on one prediction method. Different approaches get things wrong in different ways, so I run several of them next to each other — it's easier to see where they agree, where they don't, and to catch patterns a single method would miss:

• ML — an ensemble of gradient boosting models trained on historical match data.
• AI 1 — large language model #1.
• AI 2 — large language model #2.
• Math — a deterministic rule-based model built from historical patterns in the dataset.
• Books — the bookmaker market itself, represented by implied probabilities derived from pre-match odds after removing the bookmaker margin.

Books plays a different role from the other four. It doesn't predict anything on its own — it's just the market, and I keep it there as a yardstick. The bookmaker odds already bake in a lot of information, so if one of my methods can't beat that number over time, I don't think it's really earning its place.

How does it work?

Everything sits on top of a dataset of 9,000-plus T20 matches. The match, scorecard and lineup data comes from Sportmonks. Odds I pull myself, from a handful of major bookmakers, just before each game.

For every match I work out 20-odd features. Some carry more weight than others. ELO ratings do a lot of the heavy lifting, and I lean on Net Run Rate and where the two sides sit in the table. Recent form and the head-to-head record feed in as well. Then there's the lineup itself — how strong the named XI actually is, and how each team has tended to go at different stages of an innings.

The ML and AI methods learn winning patterns from all of that; Math runs the same numbers through a fixed set of rules instead.

For an upcoming game I build the exact same features beforehand, and I only run the predictions once the playing XIs and the toss are known — so nothing is guessing at a lineup that hasn't been confirmed yet.

From there each method gives both teams a win probability and picks the side it likes.

Results

Every prediction went out before the match started, and only after the toss and the XIs were confirmed.

To score them I had each method stake a flat 1,000 units on its pick. I kept the same rules for the whole season — no tweaking, no dropping the awkward results.

Transparency & track record

Each prediction is settled against the final result — a win if the pick came in, a loss if it didn't, and a no-result if the match was abandoned or never reached a decision. No-results don't count either way.

What I track is the main match prediction: who wins. Accuracy is just the share of settled picks that came in over the window you've selected; ROI% is what a flat stake would have returned at the odds that were on the board when the pick went out.

I don't delete or re-date old predictions to make the numbers look better. What you see is the whole record for the period — the losses as well as the wins — refreshed as each match finishes.

A note on variance

Even a real edge runs into short-term variance. What you see here is meant to hold up across a long run of matches — it won't tell you who wins any single game, and a good record so far is no promise of the same going forward.

Public data only

Everything I use is public — past results, the lineups teams announce, the toss, and the odds on the market. There's no insider or private data anywhere in this. Whatever edge there is comes from putting that public information together across several methods, not from seeing something nobody else can.

Responsible use

Cricombo is here for information and for fun. It isn't betting advice and it doesn't guarantee anything. If you do bet, only do it where it's legal, only if you're old enough, and only with money you're comfortable losing.