Funding & Balance Loading…

Loading…

Today

Loading…

▾Intraday PnL chart click to toggle — cumulative net PnL since 09:00 UK

Loading…

Live Operations live

Period:

Loading…

▾Recent Live Bets click to toggle

Sport: Status:

Loading…

▾Recent Live Alerts (24h) click to toggle

Loading…

📄 Paper research view — per-arm performance + gates (Phase 1.6 uses per-arm gates, not a portfolio scorecard) Loading…

Loading…

P&L for period

Period:

Loading…

Cumulative project P&L since project start

Bucket:

Loading…

Betfair Premium Charge tracker UK Exchange threshold

Loading…

Project costs

Name	Kind	Category	Date / Start	End	Amount	Accrued
Loading…

Martingale-on-favourite — paper strategy monitoring

Loading…

📖 How the system works

An end-to-end research platform that predicts sports outcomes, finds bets where the model disagrees with the Betfair market, and places real money on the Exchange — with safety guards at every step.

Right now (July 2026) the project is in the Phase 1.6 restructure window (locked 2026-07-03, review ~2026-07-31). Live capital is deliberately small — 3 arms, £464 combined, £36 in uncommitted reserve — so the biggest open question (execution deflation on limit orders) can be answered on real fills before more capital is deployed.

The five-step loop

Every bet goes through the same pipeline, the same way, every time:

Ingest — a live data feed from Betfair captures every tick (price, volume, who’s lining up, who scratched). Background loaders pull supplementary data from racing form sites, Cricsheet, Sackmann tennis data, weather APIs, etc.
Build features — for each upcoming race or match, the system constructs a snapshot of everything known at that moment: form, draw, jockey, ground, weather, recent ball-by-ball stats, head-to-head record, and so on. Zero look-ahead bias — only data that existed before the bet would have been placed.
Predict — a trained LightGBM model takes the features and outputs a probability for each runner. The probability is calibrated (e.g. when the model says “30%”, roughly 30% of those actually win).
Decide — the strategy compares the model probability against Betfair’s implied probability (1/odds). If the model thinks an outcome is significantly more likely than the market does, that’s an edge. The strategy bets only when edge ≥ a threshold (typically 5%) AND market volume is high enough AND the race is within a defined pre-off window.
Bet & settle — Kelly criterion sizes the stake from the edge (capped at 5% of bankroll). In live mode the bet goes to Betfair via API; in paper mode the bet is logged but no money moves. After the event settles, the system computes profit/loss and CLV (closing line value — did we get a better price than the market closing price?).

Paper vs Live mode

Every strategy starts in paper mode — bets are logged but no real money is risked. Phase 1.6 does not use a single “10-gate checklist” for promotion any more; that was Phase 1 vocabulary from before the 2026-07-03 restructure. Each arm now carries its own per-arm decision gate captured in the review skill and in project memory. Examples:

id=20 vs id=43 live A/B — verdict at n_paired = 100 (~2026-07-25): promote the winner, retire the loser.
id=23 GR back walker — Sortino > 0.4 rolling 30d (already cleared once at 0.43 on 2026-07-03, triggering the bankroll scale from £146 → £220).
id=11 cricket in-play — the recall gate (retrospective ball-state validation): mean recall ≥ 0.70 at n ≥ 30 evaluable matches.

Paper is directional evidence, not a promotion trigger on its own. The primary evidence source for “does the execution-deflation fix work” is the live A/B between id=20 (walker) and id=43 (limit_only) — both running on real money, same model, same bankroll, same kelly.

Cross-sport freeze: no new sport gets live capital while cricket is broken. Cricket is the portfolio’s bond — a whole-project sanity check. Live tennis will not happen until id=11’s recall gate clears and the cricket-freeze policy is released.

Regardless of any individual strategy’s mode, the box-level BETTING_MODE environment variable is the master arm: setting it to paper halts every live strategy globally.

Safety guards

Per-bet cap: 5% of bankroll, no exceptions.
Daily loss halt: currently 60% of bankroll (raised from 10% during the 2026-06-05 GR data-collection window and kept loose so per-strategy noise days don’t choke off evidence). The portfolio-level drawdown watcher does the real work here — see below.
Portfolio drawdown watcher: scripts/portfolio_dd_watch.py runs hourly, tracking peak-to-trough on deployed capital + cumulative net PnL. Warning email at 20%, breach email + auto-trip at 25% (Phase 1.6 exit Gate 5).
Micro-Kelly, not half-Kelly: the three live arms use kelly fractions of 0.03125 / 0.03125 / 0.0425 — much tinier than half-Kelly (0.125). Reason: at Phase 1 stake sizes, Betfair’s 2% commission was consuming ~69% of gross profit. We sized down to keep exposure minimal while the execution-fix experiment (id=20 vs id=43) plays out. Kelly gets scaled back up once evidence justifies it.
LAPSE persistence: every bet is a LIMIT order at the requested price; if Betfair can’t match it before the off, the unmatched portion cancels automatically (no overnight exposure).
Mode-change audit: every flip between paper and live is logged with a DB trigger. A strategy can only go live if kelly_fraction and bankroll are both set, and multi-row live flips are rejected.

Phase 1.6 exit gates

Six gates must hold at the ~2026-07-31 review for Phase 2 entry. Any miss triggers a Phase 1.7 restructure instead.

Paper screen not clearly killed — at least one of the HR paper A/B feeds (id=42 walker ctrl v2, id=44 60s cadence, id=45 30s cadence) is not showing Δ_CLV < −200 bps at n ≥ 50. Directional signal only.
Live A/B verdict at n_paired = 100 — id=20 walker vs id=43 limit_only. Promote limit_only if Δ_CLV ≥ +50 bps AND Δ_lapse ≤ +10pp. Retire it if Δ_CLV < 0 OR Δ_lapse > +25pp. Hold both to n=200 otherwise.
id=23 Sortino > 0.4 rolling 30d — the mean-daily / downside-stdev ratio replaces the old “% positive days” gate because id=23’s payoff distribution is right-skewed longshot.
Combined live PnL ≥ £0 net commission over the Phase 1.6 window (fresh window — pre-2026-07-03 PnL from now-demoted strategies is excluded).
No −25% drawdown breach at the portfolio level, per the DD watcher.
RESEARCH_LOG.md committed at least weekly — an ops-discipline gate added after a June DD event went unnoticed while the log was dark.

Per-sport detail

Click any sport to expand. Sport-specific data sources, model approach, and current strategy state.

▾ 🐕 Greyhound Racing Live · £220 · sole live arm

Greyhound racing in the UK and Ireland runs across 40+ tracks with very high frequency — over 100 races a day — which gives the model far more training data than horse racing, and the closing market is liquid enough that real-money execution doesn’t move prices. GR is the longest-running live sport in the portfolio.

Data sources

Betfair Exchange stream — live tick data on every Win market (price, volume, traders).
Racing Post API — racecards for most tracks; about 14% gap on BAGS daytime meetings.
SportingLife scraper — fills the Racing Post gap, gets us to 100% UK/IRE coverage.
GBGB historical archive — backfilled form, sectional times, trainer/dog history.

Model approach

LightGBM gradient-boosted classifier with chained isotonic calibration. Features include: trap number, recent finishing positions, sectional time pattern, days since last run, age, weight, grade, distance, track-specific historical bias. The output is a per-runner win probability that gets normalised across the field (so 6 dogs sum to 1.0).

Calibration is the strongest part of the model: across 2.6 million out-of-fold predictions the 10 deciles all hit within 0.92–1.03 of expected. Brier score 0.152, ECE 0.013. When it says “30%”, roughly 30% win.

Strategies running

Code	Mode	Notes
`gr_tier2_paper_micro_walker_floor`	Live	Sole GR live arm. £220 bankroll (scaled from £146 on 2026-07-03 21:44 UTC on Phase 1.6 Gate 3 clear — rolling-30d Sortino hit 0.43 vs the 0.40 bar), kelly held at 0.0425 because Sharpe was too low to justify concentration on a right-skewed book. Wed-skip experiment: daemon paused Wednesdays via cron since 2026-06-24.

Gate window

Bets only fire when the race is 60–180 seconds before off, market volume is ≥ £5,000, and the model’s edge over Betfair implied probability is ≥ 5%. These gates were tuned by an A/B experiment in mid-May; the “tight” arm beat the wide arm by +2,252 bps of CLV.

Why no GR lay?

GR lay was extensively researched — 7 rounds of live lay experiments over June 2026 with different target-edge, cap-band, and 2nd-favourite variants — and comprehensively falsified. Every arm was retired by 2026-07-03. No lay strategy goes live again until a paper limit-only experiment closes the observed +500 bps deflation gap between shadow paper CLV and live lay CLV. Paper twins continue to run in the background as deflation-research reference data.

▾ 🏇 Horse Racing Live · £244 · Phase 1.6 A/B

Horse racing is the centrepiece of Phase 1.6. Two live arms are running as a paired A/B on real money to answer the single biggest open question in the project: does the execution-deflation fix work? Same model, same bankroll, same kelly — the only variable is how the order is placed.

HR is the hardest sport to model because it’s feature-rich (jockey, trainer, going, draw, weight, headgear, sectional times) but lower-frequency than greyhounds (maybe 30 UK races a day) and the market is much sharper.

Data sources

Racing Post (rpscrape) — pre-race racecards plus deep historical form, 99% UK/IRE.
Betfair Exchange — live prices + Betfair Starting Price (BSP) for CLV computation.
OpenMeteo — going-relevant weather data (rainfall in the 24h before the race, etc.).
Form.aspx scraper — Australian racing data for the multi-region expansion track (paper only, id=28).

Model approach

LightGBM with the same calibration pipeline as greyhounds, but using horse-specific features: weight, official rating, draw, distance, race class, going, jockey/trainer rolling stats, head-to-head with other runners. Tier-2 model AUC ~0.66. Country whitelist locked to ["GB", "IE"] because 99% of training data is UK/IRE — AU/US/FR/ZA races are out-of-distribution and scrubbed via a SQL-level filter.

Strategies running

Code	Mode	Notes
`hr_tier2_paper_walker`	Live	id=20. £122, kelly 0.03125. The A/B control: steps the order price toward the market over time to try to get a fill. Lifetime CLV +490 bps but ~35% lapse rate — leaves a lot of edge on the table.
`hr_tier2_paper_walker_limit_only`	Live	id=43. £122, kelly 0.03125. The A/B variant: rests at fair value and never walks. Tests whether limit-only mechanics survive at a lower fill rate but zero adverse-selection drag. Flipped paper→live 2026-07-03.
`hr_tier2_paper_walker_ctrl_v2`	Paper	id=42. £25. Paper walker control on the queue-aware v2 simulator. Directional signal only.
`hr_tier2_paper_walker_ctrl_v3_60s`	Paper	id=44. £25. Cadence A/B control — 60-second poll cadence.
`hr_tier2_paper_walker_ctrl_v3_30s`	Paper	id=45. £25. Cadence A/B variant — 30-second poll cadence.
`hr_tier2_paper_walker_auus`	Paper	id=28. Multi-region AU/US variant (walker mechanic on Australian + North American racing). Paper only pending calibration + volume evidence.
`hr_tier2_paper_lay`	Paper	id=31. Paper-only HR lay experiment. Blocked from live promotion by the GR lay deflation problem.
`hr_martingale_fav_paper`	Paper	id=36. UK/IRE HR Martingale-on-favourite paper experiment. £500 paper bankroll, chain-abandon at 25 losses.

A/B verdict criteria

At n_paired = 100 (~2026-07-25):

Promote limit_only if Δ_CLV ≥ +50 bps AND Δ_lapse ≤ +10 percentage points.
Retire limit_only if Δ_CLV < 0 OR Δ_lapse > +25 percentage points.
Hold both to n=200 otherwise.

Retreat trigger: combined lifetime PnL < −£40 at n ≥ 100 → halve kelly on both arms, extend to n=200.

▾ 🏏 Cricket Paper · portfolio bond · on hold

Cricket is the portfolio’s bond. Under the cross-sport freeze policy, no new sport gets live capital until cricket unblocks. This affects tennis promotion, football planning, and anything else waiting downstream.

Recent history

cr_t20_preoff_paper (id=9) went live 2026-06-04 with a micro-stake cap. It was demoted back to paper on 2026-07-03 after a −409 bps live CLV was diagnosed as adverse-selection: matched-CLV was negative while lapsed-CLV was +2,038 bps — the same passive-limit-order signature that killed GR lay. It’s now paper for a 14-day observation window while the mechanic is re-thought.

Data sources

Cricsheet — ball-by-ball archive of every T20 international and major franchise game since 2008. ~3 million ball-rows in training. Lags 2–3 weeks (used for retrospective validation only).
Betfair Exchange — Match Odds market on T20 internationals + IPL + Big Bash + The Hundred.
Cricbuzz + Cricinfo — live ball-state scrapers. Cricsheet doesn’t stream live data, so we built sidecar scrapers for in-play current score + wickets fallen.

Model approach

Pre-match model looks at toss winner, toss decision, batting-first team, venue history, recent team form, head-to-head record, gender, format. Outputs P(team A wins).

In-play model reuses the tier-2 model with a rule-based-v2 ball-state estimator that infers current score + wickets from the live scrapers. CLV is computed against a synthetic close (LTP at scheduled off) since cricket markets don’t have BSP.

Strategies running

Code	Mode	Notes
`cr_t20_preoff_paper`	Paper	id=9. Demoted from live 2026-07-03 after adverse-selection diagnosis. £37 held in reserve. Under 14-day observation window.
`cr_t20_inplay_paper`	Paper	id=11. In-play framework running the tier-2 model + rule-based-v2 ball-state estimator. Blocked from live by the recall gate + cricket-freeze policy.

The recall gate

The in-play strategy’s promotion gate is two conditions in AND:

Retrospective ball-state recall ≥ 0.70 at n ≥ 30 evaluable matches. Currently 0.679 at n=56 — past the sample threshold, 2.1 percentage points short of the recall threshold, and stalled at that level. Populated daily by scripts/validate_cricket_ball_state.py writing to the cricket_ball_state_validation table.
Cricket-freeze policy released at the portfolio level.

Two Phase 1.6 workstreams are targeting the recall gap:

Option B (landed 2026-07-07): a runtime confidence gate on the in-play daemon that refuses bets whose ball-state inference comes from low-precision detector sources.
Option C (in flight): a Cricbuzz-precision fix targeting the 19% precision figure on ~30k detector balls.

▾ 🎾 Tennis Paper · awaiting first bets

Tennis has the cleanest research dataset of any sport thanks to Jeff Sackmann’s public archive, but the highest variance per match because it’s a binary outcome with sharp markets. Both tennis strategies are paper-only and — even if paper CLV clears the +200 bps live-promotion floor — they can’t go live while the cricket freeze is in force.

Data sources

Sackmann tennis data — every ATP/WTA match since 1968 with surface, round, draw size, result, set scores.
Betfair Exchange — Match Odds with a 60–1800 second pre-off betting window.

Model approach

Per-player feature snapshot at the time of the match: recent form (last 10 matches), surface-specific Elo rating, head-to-head record against opponent, age, fatigue (matches in last 14 days), tournament tier and round. Two perspectives (player A and player B) are predicted independently then normalised. Raw model is well-calibrated on tennis (CV ECE 2.55%); the GR isotonic wrapper did not transfer.

Strategies running

Code	Mode	Notes
`tn_atp_paper`	Paper	id=7. Main-tour ATP + WTA, 60–1800s pre-off window. Daemon died 2026-06-08 for reasons never fully diagnosed (the underlying TN-MO shard was fine; the strategy process just stopped placing bets). Revived 2026-07-06 evening; still awaiting first post-revival bets.
`tn_atp_paper_mids`	Paper	id=19. Restricted mids-only variant (odds 3.0–5.0). Same 2026-06-08 death and 2026-07-06 revival as id=7.

▾ ⚽ Football Planned

Phase 1 sport but no strategy is planned before Phase 2 at the earliest. Historical tick data continues to be ingested for the top leagues in the background so a future model has training data ready, but no live or paper strategy exists today. Football is the highest-variance per-event sport in scope (90 minutes of stochastic events, often ending in a 1-1 score) and we’re prioritising sports where the model has a cleaner edge first.

When we get to it: features will include team Elo, recent xG, lineup strength, home/away, manager tenure, fixture density.

Key terms

Edge: The gap between the model’s probability and the market’s implied probability. If the model says 30% and the odds are 4.0 (implied 25%), the edge is +5%.
CLV (Closing Line Value): Did we get a better price than the market closing price? Positive CLV is the strongest proof that the model is finding genuine information the market missed. Measured in basis points (1% = 100 bps).
BSP: Betfair Starting Price — the official close price set by Betfair for racing markets. Used as the “close” for CLV calculations. Non-BSP sports (cricket, tennis) use a synthetic close (LTP at scheduled off).
Kelly criterion: A formula that sizes each bet based on edge and bankroll. Full Kelly maximises log-wealth but has huge variance. Phase 1.6 uses micro-Kelly (0.03125–0.0425) rather than half-Kelly (0.125) to keep commission drag bounded while the execution-fix experiment runs. Kelly gets scaled back up once evidence justifies it.
Calibration: How well the model’s probabilities match reality. A perfectly calibrated “30% pick” wins 30% of the time. Measured by Brier score (lower is better) and ECE (Expected Calibration Error — lower is better).
Adverse selection: The failure mode where a passive limit order only fills when the market moves against you — matched-CLV is negative while lapsed-CLV is very positive. Diagnosed on GR lay (June 2026, 7 rounds) and on cricket pre-off (2026-07-03). The Phase 1.6 A/B (id=20 walker vs id=43 limit_only) is the live experiment to isolate the effect on HR.
Walker vs limit_only: Two order-placement mechanics. A walker steps the order price toward the market over time to increase fill probability (at the cost of some slippage). A limit_only order rests at fair value and cancels if unmatched (zero slippage, higher lapse rate). The entire Phase 1.6 A/B is about which one wins net of commission on real fills.
Sortino ratio: Mean daily PnL divided by the standard deviation of negative-only daily PnL. Used as the id=23 GR back exit gate (target > 0.4 rolling 30d) because id=23’s payoff distribution is right-skewed — a normal Sharpe ratio understates the strategy by penalising winning days.
Recall gate: For the cricket in-play state estimator: retrospective validation against Cricsheet ground truth (which lags 2–3 weeks). Currently 0.679 at n=56, target 0.70 at n≥30. A complementary runtime confidence gate landed 2026-07-07 that refuses bets from low-precision detector sources.
LAPSE persistence: A Betfair order type. The bet sits at the requested price until the market starts; whatever didn’t match by then is cancelled, never left dangling.

Portfolio Progress — Plan toward £100/day milestone

Current phase

—

Long-run milestone

—

Per-arm gate model (Phase 1.6)

Portfolio drawdown vs ceiling

—

/ ceiling —

Live bankroll

—

Cumulative net PnL

—

Cricket recall gate: loading…

Phase 1.6 exit gates —

—

Live arms — informational 14d rolling stats (per-arm gates drive Phase 1.6, not this table)

Strategy	Sport	Bankroll	14d mean £/day	% pos days	Worst day (% bnk)	Sharpe	Cum PnL £ (net)	Status	Scale step
Loading…

Paper arms — Phase 1.6 A/B pipeline + demoted arms with pending decisions

Paper arms feeding Phase 1.6 exit gates (id=42/44/45 A/B pipeline) plus demoted arms with pending decisions. Cum PnL below = LIVE PnL since Phase 1.6 start (2026-07-03) — non-zero only if the arm was still live briefly inside the window. Paper PnL sits in the per-sport reviews (/hrlivetest, /grlivetest, /crlivetest).

Strategy	Sport	Bankroll	Pre-demote live PnL £ (in plan window)	Next decision	Decision date
Loading…

Upcoming pre-committed decisions

Date (UTC)	Strategy	Decision
Loading…

Add cost

📖 How the system works

The five-step loop

Paper vs Live mode

Safety guards

Phase 1.6 exit gates

Per-sport detail

Data sources

Model approach

Strategies running

Gate window

Why no GR lay?

Data sources

Model approach

Strategies running

A/B verdict criteria

Recent history

Data sources

Model approach

Strategies running

The recall gate

Data sources

Model approach

Strategies running

Key terms

Live arms — informational 14d rolling stats (per-arm gates drive Phase 1.6, not this table)

Paper arms — Phase 1.6 A/B pipeline + demoted arms with pending decisions

Upcoming pre-committed decisions