Locked markets
Win / O/U / BTTS are model-binding. The LLM cannot rewrite them.
Most “AI tips” are a prompt plus one price.
We run a reproducible inference pipeline. Every match gets the same feature pack, the same blend, the same decision layer. Copy comes from the LLM – binding tips come from the model. When book, API-ML and Dixon-Coles disagree, confidence drops. When they agree, it rises. That’s research, not a newsletter.
Consensus odds
Median across Bet365, Bwin, Unibet, 1xBet, Betclic – not the first hit.
Compute budget
Top leagues first. U21/B sides last. 70% of refresh slots for tier A.
What BetKings24 is not
- Not a bookmaker – we do not take stakes.
- No brokerage, no affiliate links to operators.
- No profit guarantee. Predictions are non-binding.
Three planes, one decision
Data, inference and policy are separate. The model stays honest even when the prose is fluent.
01 · Data plane
Fixtures, predictions, squads, scorers, XI, injuries, H2H, last-5 including xG, venue split, rest days, table, referee. 28 competitions, calendar window, early-cup minnow filter.
02 · Model plane
Dixon-Coles λ with league prior and shrinkage, bookie de-vig, API-ML. Blend weights by sample size. O/U line 1.5 / 2.5 / 3.5 by λ edge. Confidence from the probability gap.
03 · Policy plane
Top-pick scoring, auto-publish, draft the rest. Combo builder with correlation caps, joint-P and EV gate. Same-game parlay. Settlement feeds calibration.
Pipeline v11 – fixture to settlement
One run, twelve stages. Nothing is skipped – it is prioritised.
01
Ingest
28 competitions, calendar window
Top-5, CL/EL, other top flights, second tiers, cups. Early amateur rounds are dropped before they burn the analyse limit.
02
Queue
League tier before kickoff time
Tier A (top-5/EL/CL) first, U21/B/Jong/II last. The limit is for quality, not whoever kicks off earlier.
03
Features
The full match vector
API predictions, median odds, squads, top scorers, assist/card boards, XI, injuries, H2H, form, last-5, venue split, rest days, table, league goal averages.
04
λ model
Dixon-Coles, not Elo folklore
Attack/defence with league prior and shrinkage. Missing stars and XI compress λ. Rest days adjust attack/concede. Last-5 home/away splits the base.
05
Ensemble
Three engines, one blend
Bookie de-vig × API-ML × Dixon-Coles. Early season leans market, later leans model. Source dissent cuts confidence, consensus lifts it.
06
Decision
1X2 · line · BTTS
AH line steers the DRAW threshold. O/U picks 1.5 / 2.5 / 3.5 by modelled edge, not always 2.5. BTTS from the same matrix.
07
Confidence
Gap + calibration + league
Raw conf from p(tip) plus gap to the second side. Isotonic calibration on 120 days of published win tips. League hit rate (n≥8) adjusts ±2–4 points.
08
LLM lock
Prose yes, override no
The LLM model writes analysis, exact score, scorers, corners/cards. sieg.tip, over_under.tip, beide_treffen.tip are taken from the blend and hard-overwritten.
09
Refresh
0.4–16 h before kickoff
Odds, XI, injuries again. 70% of slots for tier A. Model patch without a full LLM pass; re-analyse only on tip change, official XI or injury delta.
10
Policy
Only top picks go live
Score from conf, odds sweet-spot, value, source consensus, league HR. Free is strict, Pro wider. The rest return to draft – including combos that contain drafts.
11
Combos
Banker · 3-fold · same game
Upcoming NS only. Max one Over 2.5, one BTTS Yes, two games per league. Joint-P and EV gate. 4-folds Pro only. Same game: two markets, one match, no contradiction.
12
Loop
Settlement → calibration
FT results settle tips and combos. Hit rates by market/league and conf bins feed the next run. The system learns from what was live – not from drafts.
Ensemble weights by sample size
The fewer games the sides have played, the more we trust the market. That is shrinkage, not vibes.
Early · n<6
55% Bookie de-vig30% API-ML15% Dixon-Coles
Mid · n<12
45% Bookie de-vig33% API-ML22% Dixon-Coles
Stable · n≥12
40% Bookie de-vig30% API-ML30% Dixon-Coles
p = wbk · Bookiedevig + wapi · API-ML + wdc · Dixon-Colesλ,ρ
Confidence is not a marketing number
A 70 conf does not mean “it will happen”. It means: after gap, calibration and league hit rate, modelled hit probability sits in that bin. DRAW stays hard-capped (48–62) because draws are the expensive 1X2 error.
Probability gap
Bonus only on clear separation (gap > 8 pp). Tight games stay sober.
Isotonic bins
120-day calibration on published win tips. Too little sample → raw conf.
League hit rate
n≥8: strong leagues +2, weak −2/−4. Free lockout under 45%.
The LLM is an analyst, not the referee
Many sites let ChatGPT invent the pick. Here the model sees the feature pack, the LLM writes the rationale – and apply_model_locks overwrites 1X2, O/U and BTTS with the blend. Corners/cards only if last-5 match stats exist on both sides with n≥3. Exact score and scorers stay LLM markets with lower priority.
Policy layer: what is allowed live
Not every analysed draft is a tip. The policy maximises long-run expected value under uncertainty – the RL idea, without pretending we trained a neural net overnight.
Top picks
Free: conf, odds window, no LLM error, no weak league. Pro: wider, but no 46 conf.
Banker 2-fold
Two high-conf legs, odds ≥1.20. Short favourites as 1X, Over 1.5 as the safety leg.
Same game
Two markets, one match. Allowed: 1X2/DC + O/U or BTTS. Forbidden: Over+Under, BTTS-Yes+Under 1.5.
Correlation
Max one Over 2.5, one BTTS Yes, two games in the same league. Joint-P with penalty, then EV gate.
Information that arrives after the first pass
Monday’s tip is worthless on Saturday if the striker is out. So we refresh 0.4–16 h before kickoff: odds, official XI, injuries. 70% of slots for tier A. U21 games do not burn budget ahead of Inter–Udinese.
What other sites don’t do
| Typical tip blog | BetKings24 pipeline v11 | |
|---|---|---|
| Source | One prompt, one screenshot | API databases + 5-book median |
| Model | “The AI says 1” | Dixon-Coles × de-vig × API-ML, weights by n |
| Who decides | LLM can override anything | Hard lock on 1X2 / O/U / BTTS |
| Draws | Ignored or guessed | AH threshold, conf cap 48–62, own bin |
| Lines | Always Over 2.5 | 1.5 / 2.5 / 3.5 by λ edge |
| Freshness | One run the day before | Pre-KO refresh, XI and injury delta |
| Selection | Post everything | Top-pick policy, rest stay draft |
| Combos | Stack three favourites | Correlation, joint-P, same game, 4-folds Pro only |
| Learning | No feedback loop | Settlement → conf bins → next run |
The stack we actually run
API databasesDixon-Coles ρBookie de-vigAPI-ML blendIsotonic conf binsLeague hit rateLLM lockPre-KO refreshEV policySame-game parlay
Competitions covered
Access
Try Pro for 7 days – no credit card, no auto-renew. Access ends by itself. After that, a one-time payment for 1, 3 or 6 months. No forced subscription.
0 € · 7-day trial
€8.99 · 1 mo
€22.99 · 3 mo
€34.99 · 6 mo
Not an invitation to gamble. Predictions are non-binding, we are not a bookmaker, 18+. Past results do not indicate future hit rates. · Guide