# Betting For A Living Model Backbone

## Conviction Backbone V2

The public regular-football model is now market-router first.

Automatic publication remains heavily restricted while Asian handicap is promoted into the normal importer and paper-trading loop. 1X2 is still collected and stored for research, but it is not auto-published in the regular-football lane without a longshot catalyst and a matching market expression.

## Principles

- Market prior first. The model starts from bookmaker consensus, not from a narrative.
- Historical calibration second. Odds buckets are calibrated against imported football-data history.
- News is an overlay. Injuries, illness and suspensions can move a price, but they cannot replace the market prior.
- Long 1X2 prices are treated as favourite-longshot-bias risk.
- Away wins need extra proof.
- Longshots need a catalyst. A price gap alone is not a bet.
- Do not fade a strong-name side for a rough patch unless the problem is systemic and confirmed by news, squad availability, line movement or repeated performance decline.
- Once a paper tip is public, it stays in the record.

## Live Regular-Football Flow

1. Pull strict launch-scope fixtures and odds.
2. Store 1X2, over/under 2.5 and Asian handicap quotes.
3. Build de-vigged bookmaker consensus.
4. Apply historical odds-bucket calibration.
5. Apply capped team-news memory overlay.
6. Route the thesis to the right market: Asian handicap/DNB first, totals second, 1X2 only with a catalyst.
7. Publish only when edge, confidence and guardrails pass.

## Guardrails

- Minimum 1X2 edge: 5.5 percentage points.
- Minimum totals edge: 3.5 percentage points.
- Longshot threshold: 3.75+.
- Longshot minimum edge: 7.5 percentage points.
- Longshot catalyst gate: no automatic longshot publication without a verified injury/suspension/illness shock, manager disruption, fixture congestion/rotation, market move from opener, or a multi-match systemic decline.
- Strong-name rough-patch rule: do not fade an established strong side solely because of short-term form. Prefer draw-no-bet or Asian handicap if the price is awkward.
- Away-win minimum edge: 7.0 percentage points.
- Away wins above 3.25 are blocked from automatic publication.
- Maximum news overlay movement: 2.5 percentage points.

## Longshot Gate

The early 2026-05-11 regular-football picks exposed a bad failure mode: the model accepted long draws and an away outsider because the available price looked above consensus. That is not enough.

From now on, long 1X2 selections are research-only unless there is a named catalyst and a tradable market confirmation. Acceptable catalysts include:

- verified injury, illness or suspension to a key player or unit;
- credible rotation or fixture-congestion pressure;
- manager disruption or structural tactical change;
- strong opening-to-current price movement in the same direction;
- repeated underlying weakness across several matches, not just a famous team having a rough patch.

When the signal is really "the favourite is vulnerable", the preferred expression is Asian handicap, draw-no-bet or totals. The model should not jump straight to draw or away longshot.

The Tottenham v Leeds paper trade on 2026-05-11 is the example. Leeds outright at 4.50 lost in a 1-1 draw, while Leeds +0.5 would have won and either side draw-no-bet / Asian 0 would have pushed. The view needed a handicap expression, not a 1X2 outsider.

## Backtest Summary

Latest holdout backtest:

- Train before: 2022-07-01
- Test from: 2022-07-01
- Matches tested: 21,277
- Best-price total picks: 61
- Totals ROI: +19.8%
- Average-price picks: 0

This is not enough to claim an edge. It is enough to justify the operating rule: publish rarely, prefer totals, and avoid long 1X2 unless a future backtest proves otherwise.

## Next Research

Football-Data does include Asian handicap odds (`AHh`, `AHCh`, `AvgAHH`, `AvgAHA`, `MaxAHH`, `MaxAHA`, `AvgCAHH`, `AvgCAHA`, `MaxCAHH`, `MaxCAHA`). The first standalone AH report is now published at `/asian-handicap-backtest.md`.

Asian handicap looks like the right next research lane because it avoids a lot of the draw and longshot favourite bias in 1X2. It should not be auto-published until the importer stores AH fields, daily snapshots are paper-traded, and the live record shows the prices were actually tradable.

## Challenger Models

The challenger run after the 2026-05-11 failure is published at `/articles/challenger-model-backtest.html`.

Read from that run:

- Naive 1X2 price gap is rejected despite a positive tiny historical sample.
- Strong favourites are not a standalone edge.
- Simple totals buckets are cleaner than 1X2 but too thin by themselves.
- Asian handicap line movement is the strongest research signal, but it must be made live by storing openers and current lines.

## Serious Model Lab

The overnight serious model lab is published at `/articles/serious-model-lab.html`.

The blunt result: no standalone live model beats the market reliably enough to become an auto-tip engine. The next backbone stays as a market router:

- store opener immediately for 1X2, totals and Asian handicap;
- use Poisson/Elo as capped overlays, not as market-overruling brains;
- route favourite-vulnerability views to Asian handicap or draw-no-bet before 1X2;
- keep the longshot catalyst gate;
- promote Asian handicap movement only after live opener/current-line paper trading proves it.

Market baseline from the lab:

- Opening 1X2 log loss: 0.9991
- Closing 1X2 log loss: 0.996