🛡 Closing-line guarantee:🎁 Referral:

Proba Bet
Try free

🧠 How Babeth learns

An algorithm that learns from its mistakes, every night

You win about one bet in two whatever you do: the result alone teaches nothing. Babeth learns from what matters, the closing line, and only changes method when the new one does clearly better on bets it has never seen.

The model in charge

in charge

Decision trees · learned guardrails · since November 1

Hundreds of small decision trees that spot the combinations that beat the close (bookmaker × sport × Pinnacle movement…).

106
bets a day in the selection
+4.2 %
average closing line value (CLV)
75 %
beat the closing line
+44 %
expected gain a day versus the previous rule

Judged from October 19 to November 1 on bets it had never seen. Expected gain = suggested stake × closing line value, the dashboard's rule.

Every night, in three steps

1

Replay the past without cheating

For each day of the last five weeks, every model learns from the bets whose match had already started, then predicts that day's bets. It never sees the answer before predicting.

2

Judge on the closing line

A mistake isn't a lost bet: it's a price the close has beaten. Each candidate is replayed over 14 days with the engine's selection rules (except the two-scan confirmation, which can't be replayed) and scored on what it would have earned at the close.

3

Only change for the better

The model in charge is only replaced by a candidate that is clearly better day after day (statistical test), with no lower closing line value and no over-optimistic predictions. It is then retrained every 6 hours.

Last night's candidates

Same bets, same 14 days, same selection rule: only the method changes.

  • ✅ Decision trees · learned guardrails
    106.2 bets/day · CLV +4.15 % · 75 % beat the close · Expected gain +44 %
  • Trees + match factors · learned guardrails
    101.4 bets/day · CLV +4.13 % · 74 % beat the close · Expected gain +35 %
  • Additive + trees blend · learned guardrails
    84.1 bets/day · CLV +4.50 % · 75 % beat the close · Expected gain +33 %
  • Blend + match factors · learned guardrails
    84.6 bets/day · CLV +4.47 % · 74 % beat the close · Expected gain +29 %
  • Additive model · learned guardrails
    67.2 bets/day · CLV +3.93 % · 72 % beat the close · Expected gain +11 %
  • Additive model · fixed guardrails
    54.4 bets/day · CLV +4.18 % · 73 % beat the close · Expected gain 0 %

Expected gain: difference with the previous rule (additive model, fixed guardrails) on the same days.

Learned guardrails

Some rules used to rule bets out automatically: handicaps, leagues deemed unreliable, long odds, small displayed edge… Babeth opens them sport by sport when, over 21 days, the bets she would have kept despite them clearly beat the close, and closes them again as soon as they no longer do. A sport that doesn't yet have enough bets to be judged on its own follows the evidence from all sports combined. Nothing is opened without evidence.

The bet review

Every settled bet in the selection is reviewed. Losing at a good price isn't a mistake: it's bad luck, and it fades over time. A price beaten by the close is one: Babeth looks for why.

Last 30 days: 2,928 settled bets

Well played 34 %
good price, won
Bad luck 33 %
good price, lost
Lucky 15 %
price beaten, won anyway
Mistake 19 %
price beaten, lost

67 % of these bets had a good price (above the fair closing odds). Over 7 days: 75 % (718 bets).

Why the close beat the price

  • Immediate correction: Pinnacle moved to the bookmaker within half an hour (the bookmaker knew something)47
  • Last hour and a half: line-ups, absences, pre-match news92
  • Gradual market drift242
  • Pinnacle didn't move: the signal's reference was too optimistic40

⚠️ Where Babeth gets it wrong most

  • Football · totals · over 24 h before147 bets · CLV +0.2 %
  • Football · totals · 1 to 6 h before72 bets · CLV +0.5 %
  • Tennis · match result · 6 to 24 h before76 bets · CLV +1.1 %

✅ Where she does best

  • Ice hockey · totals · over 24 h before43 bets · CLV +10.3 %
  • Ice hockey · totals · 6 to 24 h before50 bets · CLV +6.5 %
  • Basketball · match result · over 24 h before24 bets · CLV +6.4 %

Babeth's log

Match factors: being collected

Since 25 October 2026, every bet is recorded with its factors. They only count in the selection if a model that uses them wins on bets it has never seen: it takes about two months of data to know.

%
BabethI only keep what works on bets I've never seen. The rest I write down, and I try again the next night.

Frequently asked questions

Can Babeth predict results?

No. The chance of winning shown comes from the market (Pinnacle, Betfair), which is already very accurate. What Babeth predicts is the value of a price: will it stay above the fair odds until kick-off? That's what pays over time.

Why judge on the closing line rather than on profits?

Because about one bet in two loses, even at an excellent price: it takes thousands of bets for profits to tell the truth. The closing line tells it from kick-off, bet by bet.

What about line-ups, weather, injuries?

Pinnacle already prices them in, very fast. Babeth records them with every bet and tests them every night: they will only enter the selection if they improve results on bets never seen before.

What if a new model gets it wrong?

It is never put in charge without doing clearly better over 14 days it had never seen. If the model in charge degrades, Babeth pauses it and goes back to the previous rule, automatically.

Sources

Odds: Pinnacle and around forty bookmakers (The Odds API), Betfair Exchange. Football matches (line-ups, absences, tables): API-Football. Weather: MET Norway (Norwegian Meteorological Institute), CC BY 4.0 licence. Stadium cities: © OpenStreetMap contributors.

The bets kept by the model in charge

In your dashboard, every bet has its “Why this bet?”, factor by factor.

See today's betsAll results