⚽ Tomulator
Overview
◈Dashboard📅Today's games◎State of play
Money
💷Bets & strategies
Explore
⊞Leagues & seasons📐Metrics▤Data sources
Engine room
🎲Simulations🧪Lab≣Catalog?How this works

strategy & model leaderboard

← lab · 24 backtests ranked by skill vs the market · ran 7/14/2026

What this is: Ranks every backtested model by skill against the bookmaker's closing prices — the fairest single scoreboard we have.
Ranked by skill first (Brier edge vs the de-vigged market — positive means sharper than the bookmaker), then CLV, then ROI. Note: entries differ in markets/periods; same-competition rows are the fair comparisons. Armed strategies await their forward paper trials.
BacktestModelCompBetsROICLVBrier edge
#29ml_gbm_v1PPL3,581-3.8%0.68%-0.00063
#28ml_gbm_v1DED4,313-1.8%-0.14%-0.00207
#27ml_gbm_v1FL15,766-5.1%0.10%-0.00221
#26ml_gbm_v1BL15,468-4.4%0.49%-0.00237
#25ml_gbm_v1SA6,574-7.2%0.52%-0.00243
#22ml_gbm_v1PD6,843-7.3%0.15%-0.00318
#10ml_gbm_v1EPL336-4.7%0.28%-0.00357
#21ml_gbm_v1EPL8,430-3.3%0.40%-0.00391
#18ml_gbm_v1EPL9,586-1.5%0.53%-0.00465
#1poisson_twEPL353-5.8%-1.05%-0.00526
#19ml_gbm_v1EPL9,793-1.9%0.38%-0.00563
#16ml_gbm_v1EPL9,588-2.9%0.42%-0.00575
#12poisson_twEPL1,442-11.4%-0.44%-0.00595
#17poisson_twEPL1,442-11.4%-0.44%-0.00595
#2poisson_twEPL2,322-7.0%-0.25%-0.00617
#20ml_gbm_v1EPL9,900-2.4%0.40%-0.00637
#14ml_gbm_v1EPL9,248-2.2%0.31%-0.00702
#15ml_gbm_v1EPL9,659-1.0%0.31%-0.00718
#13ml_gbm_v1EPL9,247-1.4%0.42%-0.00764
#7ml_gbm_v1EPL8,235-1.3%0.28%-0.00875
#11ml_gbm_v1EPL8,203-2.7%0.42%-0.00966
#6ml_gbm_v1EL29,818-5.5%-0.04%-0.01067
#4ml_gbm_v1EPL10,506-2.1%0.07%-0.02901
#3ml_gbm_v1EPL10,616-3.0%0.06%-0.03131

Registered strategies

CodeStatusDescription
draw_family_v1dormantThe sweep survivor: back the DRAW when model edge > 10% in high-tempo matchups. Verdict window +12.6% on 324 unseen bets (upper bound). [BLOCKED: model 'gbm' needs a forward-serving job the executor doesn't have yet — review finding #12; was mislabeled 'armed'.]
draw_calm_books_v1dormantSibling survivor: back the DRAW when edge > 15% and bookmaker margin is low. Verdict +4.5% on 429 unseen bets. [BLOCKED: model 'gbm' needs a forward-serving job the executor doesn't have yet — review finding #12; was mislabeled 'armed'.]
mlb_dog_variance_v1liveBaseball is the highest-variance major sport (rating->win only +0.12) yet the public loves favorites. Back MLB underdogs priced 2.2-3.5 when our Elo edge is at worst -3%.
elo_value_mlb_v1liveStraight Elo value on MLB moneylines: bet either side when Elo win probability beats the best price by 4%.
drought_fade_v1liveLaw-derived (winless-run replicates in 4 sports): bet AGAINST any side whose opponent... no — whose own winless run is 4+ games, when priced under 2.8. Fade the drought.
steam_confusion_v1liveMarket-structure play: when bookmakers DISAGREE hard on a price (cross-book spread > 8%), take the best available price on the side with any positive Elo edge — someone is wrong, take the generous one.
fatigue_fade_nba_v1armedLaw-derived (NBA congestion is real, r=-0.067): fade NBA teams playing their 4th+ game in 7 days against fresher opponents.
hot_hand_fade_v1liveMean-reversion creative: fade sides whose Elo momentum is scorching (opponent momentum in top band) — the market overprices streaks. Expected to be humbled; that is the point of testing it.
chaos_parlay_v1liveThe fun one: a $2 three-leg parlay of the day's highest-edge independent legs across all live strategies' candidates. Margin math says it should bleed; it exists to make the math visible.
placebo_rain_unders_v1

Reading the columnswhat each number actually means

ROIprofit per unit staked at opening prices
CLVhow much better our price was than the closing price — the earliest sign of real skill
Brier edgeprobability accuracy vs the de-vigged closing market; positive = sharper than the bookmaker
Spec · the reproducible recipe
{
  "kind": "leaderboard"
}
dormant
CONTROL: bet football unders when heavy rain is forecast. Our own discovery says rain does NOT reduce goals — so this placebo SHOULD lose at the margin rate. If it wins long-term, our discovery engine has a bug. Science. [BLOCKED: model 'gbm' needs a forward-serving job the executor doesn't have yet — review finding #12; was mislabeled 'armed'.]
cards_closeness_v1dormantDiscovery-derived (closeness->fouls, r=+0.19 emergent): cards/fouls overs in evenly-matched games. Dormant until a live cards-odds feed exists. UPDATE 2026-07-03: referee identity alone swings P(4+ cards) from 33.6% to 54.0% (quartile split, n=21,870) — add match.ref_cards_avg regime when the odds feed exists.
value_all_v1retiredBaseline engine: any market, 5% edge, fractional Kelly. Backtest verdict: -2.05% over 20 EPL seasons. Kept as the reference corpse.
wc_sim_value_v1liveWorld Cup: back any 1x2 selection where the Poisson fixtures-sim beats the best live price by 8%. Caveat honored: the sim runs hot on tournament data, hence the high bar. [2026-07-07: gap-filter gate mandatory — naked longshot gaps went 8W-49L in the cup retro and the gf twin vetoed them all.]
wc_upset_hunter_v1liveWorld Cup knockout chaos: longshots priced 4-10 where the sim still sees positive edge. Small stakes, big stories.
wc_ko_unders_v1liveKnockout football is cagey (legs tighten when losing means going home): under 2.5 goals when the sim agrees (any positive edge). Hypothesis-tagged.
steam_chaser_v1liveLine-movement study (78k selections, 20yr): big steam (4+ prob-pt moves) is the one bucket the close doesn't fully absorb (+2.3pts residual), concentrated in soft leagues (+5.9% EL2) and away sides (+6.4%). Hypothesis-grade (CI spans zero) — this forward trial is the judge. Bets any selection whose consensus implied prob rose 4+ pts across our own daily snapshots.
portfolio_v1liveLayer 3, the bet-maker: builds the best BASKET across everything priced (ballast/value/longshot buckets, one bet per contest, quarter-Kelly, model blended 50/50 with the market) and Monte Carlos each slate before betting. Historical replay verdict: with the current football model's edge it would have gone bust over 21 seasons ($1000 -> $0.76) — this forward trial tests whether the newer, humbler probabilities change that.
pitcher_edge_v1liveWho's on the mound matters most in baseball. Bet an MLB moneyline side when its probable starter's season ERA is at least 1.00 better than the opponent's (mlb.starter_era_gap, signed to the bet side) AND our Elo sees any positive edge at the best live price. First strategy to use contest-scoped context metrics.
prop_scorer_v1liveThe first player-level strategy: a gradient-boosted scorer model (rolling goals/shots per 90, starting rate, team attack vs opponent defence, confirmed lineups when available) prices every Kalshi anytime-goalscorer market and bets where its probability beats the fee-adjusted ask by 10%+. Settles from Kalshi's own results.
ref_cards_draw_v1livePromoted from the matrix tournament (won ALL 4 backtest seasons, worst season +$471): when the referee's card average is in the top tercile (>=3.52/game), the match tightens — back the DRAW blind. Model-free context bet; forward trial to survive multiple-comparisons doubt.
cards_over_ref_v1liveThe referee signal finally bettable: card-happy ref (top tercile, >=3.52/game) -> back total cards OVER at the main line from the new odds feed. Settles from post-match team stats.
corners_style_over_v1liveTwo corner-hungry teams meet (style clash top tercile, >=11.4 combined corner averages) -> back total corners OVER at the main line. Model-free; the relational miner showed corners follow style and mismatch.
wc_sim_value_kor_v1liveStaking A/B twin of wc_sim_value_v1: identical signal (Poisson sim beats best price by 8%) but sized by kelly_of_ruin — half-Kelly scaled by the posterior probability this bankroll's edge is real, from its own settled record. Measures the staking policy, not the signal. [2026-07-07: gap-filter gate mandatory — naked longshot gaps went 8W-49L in the cup retro and the gf twin vetoed them all.]
mls_sim_value_v1liveMLS forward trial: back any 1x2 selection where the fixtures sim beats the best price by 10%. Higher bar than WC — the sim is new to this league and MLS draws are notoriously frequent.
wc_sim_value_gf_v1retiredGap-filtered twin of wc_sim_value_v1: identical sim signal, but each edge must ALSO be blessed by the gap-survival classifier (P(model side right) >= 0.55, trained on 21 seasons of model-vs-market disagreements). The meta-filter's live A/B.
dutch_cover_v1liveHalf-cover dutch: back the sim's 1x2 pick (edge >= 5%) AND the draw, draw stake sized to refund HALF the outlay on a draw. The 21-season backtest (exp #203) found full cover overpays for insurance (median $898) while half-cover beats straight outright: median $1,060 vs $995, worst season $87 vs $17.
dd_live_v1liveIn-play double-down: when a pending pre-match 1x2 single is TRAILING at a poll and the conditioned model posterior still shows >=8% edge at the real live price, add one $10 bet on the same side. The one in-play behavior that beat its control in the 3,454-match bayes replay (+8.1% vs +2.5%, exp #206). Placed by the inplay watcher, not the executor; graded against REAL live odds.
simple_parlay_v1liveSimple 2-leg parlay of the day's PROVEN-class bets only (ko unders, ref-gated cards over, style-gated corners over — the classes that are 11-for-13 live). Two +EV legs compound edge multiplicatively; small flat stake because variance compounds too. Legs come from bets the proven strategies actually placed today.
epl_mid_dog_v1liveThe tuning sweep's best honest cell, armed for 2026-27: EPL home mid-dogs (price 2.6-4.5) when the engine's edge >= 8% -> +18% ROI over 540 real-priced bets, era-stable (+15%/+12%/+10% across 05-10/17-21/22-26). Half-Kelly per the matrix law. HONEST CAVEAT: thresholds chosen in-sample; this live trial is the out-of-sample test.
epl_mid_dog_v2livev1 + the loss-pattern filters baked in: real favourite only (away rank <= 14), our dog not in the drop zone (home rank <= 17), home side alive (form >= 4 pts l5), and no Aug-Oct bets (early-season inputs are noise). Holdout 2016-26: +31.0% over 128 bets vs v1's +15.0% on the same seasons; filters do NOT travel to other leagues (-6.7%) — this is an EPL edge. A/B twin of v1, ~13 bets/season.
prop_scorer_v2liveAnytime-scorer v2.1: involvement-first gates (forwards gi>=0.40/90, attacking mids gi>=0.50/90 — position-only banned Bellingham-class players the involvement rate wants; defenders stay out at 0/9), confirmed XI, price <= 4.5, full scorer model, edge >= 10%. Retested on the real graded window: rejects all five of v1's losers, admits nothing that lost. A/B twin of v1.
prop_soa_v1liveScore-or-assist singles with the context gates baked in from day one (graded WC window: naked SoA -60%, elite lane -51%, price 5+ 0-for-13 -> tails banned): goal involvement >= 0.40/90 over last 12 apps, price 2.0-4.5 only, expected starter, elo not against (within 50), and the keeper gate — only bet INTO a cold opposing keeper (team.keeper_form_v1 below pool median). Kalshi SoA settles via settle-kalshi-props.
soa_duo_parlay_v1liveTom's both-teams structure: when EACH side of one match has a qualifying prop_soa_v1 player, parlay the pair ($3). Mildly negatively correlated legs (one side dominating starves the other) — the trial measures whether the price makes up for it.
fade_goal_live_v1liveIn-play fade-the-goal: when a goal goes in, back the CONCEDING side at the real live price if the conditioned posterior still shows >=10% edge. The bayes replay's best surviving policy (+11.7% ROI over 1,324 bets on the own-goal-corrected sample, 6/10 winning seasons vs +1.1% control). Placed by the watcher, max 2 per contest, savepoint-safe. The definitive grade is this live sample at real odds.
mad_game_live_v1liveMad-game mode, live: the whole $1000 game budget committed to ONE match. Plan: spray_and_pray — equal-revenue dutch with a +10% return target, the one-game-format champion of the 12-plan mad-game replay (exp #225: median $1,100, 75.6% hit rate, 0.4% bust vs tranche_ladder's $758/10.7%). By 12' the watcher covers the largest outcome set whose dutch return clears the target at real live prices, stakes ~1/price (small on longshots, big on favourites), and holds the unstaked remainder as the designed loss buffer. tranche_ladder remains available via bet_rule.plan. One contest per day.
el2_mid_dog_v1retiredLeague Two home mid-dogs in the 3.5-4.5 price band when the sim engine's edge >= 8% -> +16% ROI over 161 real-priced bets (2017-25), positive in the last 5 straight seasons (+10/+20/+7/+5/+4%), one COVID-empty-stadium down year. The segmentation finder showed the edge concentrates in 3.5-4.5, not the wider 2.6-4.5 band (which was diluted to +3.5%). Half-Kelly per the matrix law (its honest-staking cells on League Two average ~+20%/season and never bust). HONEST CAVEAT: band + threshold chosen in-sample on the baseline poisson engine; this live trial is the out-of-sample test, and the tuned engine hasn't been applied yet.
goal_bets_v1livePre-match goal-market value: price total-goals (O/U 2.5) and BTTS off the Dixon-Coles score matrix built from each contest's stored sim lambdas (home/away-split, shrunk), bet where the book is soft by >=8%. Also surfaces the two likeliest correct scores. Placed by the goal-bets command, graded from the scoreline. Tom's goal-bets idea; the score matrix also feeds the half-time Bayesian re-bet (goal_bets_ht_v2).
goal_bets_ht_v2liveBayesian half-time re-bet on total goals: from ~40' the live score is data, so the score matrix conditions on it (remaining goals ~ Poisson on the remaining lambda share) and the final O/U 2.5 posterior updates exactly. If it beats the live over/under price by >=8%, back it. Placed by the in-play watcher at real live odds, one re-bet per contest. v2 of goal_bets_v1; debuts on the WC quarter-finals.
goal_bets_late_over_v3liveLate-overs goal variant: from 50' back OVER 2.5 only when the score-matrix posterior beats the live over price by >=8%. Top of the goal-replay family (exp: +11.6% ROI, 8/10 winning seasons over 1225 corpus bets) BUT on synthesised live lines where late-over longshots may be flattered by the vig model — this WC trial is the real test. Backing late UNDERS was a -20.7% trap in the same backtest, so this variant is overs-only. Watcher-placed, one bet per contest.
away_fav_continental_v1liveedge_hunt finding (2026-07-09): back the AWAY side at short/even odds (<=2.6) in Serie A / Ligue 1 / Primeira. +3.0% ROI at consensus market price, +4.7% best-of-book over 3,366 bets (2015+). Survives a temporal train/test split, BEATS the closing line, positive in 13/14 recent seasons, holds at consensus (not a best-of-book artifact; ~3 books/game). Maps to the home-favouritism bias. CAVEATS: small edge (~3%); ~2012 regime change means exploitable-now not timeless; this live paper trial is the out-of-sample test.
wc_form_edge_v1liveWorld Cup: back any 1x2 pick where the fixtures-sim beats the best live price by 5%+ AND the bookmakers disagree with each other by 4%+ on that price — tournament markets get confused, and confusion is where a model edge is most believable. The lower edge bar than wc_sim_value_v1 is paid for by the confusion gate.
wc_keeper_wall_v1liveWorld Cup club-form carryover: a keeper doesn't forget how to save when he swaps his club shirt for his country's. Back a team (1x2) when THEIR keeper's rolling save% — built from ALL his matches, club included — is hot (>=75%) and the OPPONENT'S keeper is leaky (<=68%), requiring only a non-negative sim edge so the keeper signal is the driver. NOTE: fires only once nations carry team.keeper_form values (needs AF starter data for national-team matches — the daily catchup is filling this); silent until then by design, never guessing.
wc_parlay_v1liveWorld Cup accumulator: combines today's two best singles actually placed by the WC strategies (wc_sim_value_v1 / wc_upset_hunter_v1 / wc_form_edge_v1 — every leg already passed its own gates) into one 2-leg parlay. Small stake, tournament-sized price.