FightNight
UFC rating engine + fight prediction system. Glicko-2 spine, two-stage
prediction model, market benchmarking, and closing-line-value (CLV) evaluation
against BestFightOdds opening lines.
Headline results
The model moves betting lines. Pure model picks taken at the opening line
show +1.9 to +2.0 pts of vig-free CLV in every era since 2018 (t ≈ 9–11,
n = 5,982 out-of-sample fights). Each 10 pts of model-vs-opener disagreement
predicts ~2.2 pts of line movement toward the model by close. Regime break at
2018 coincides with PASPA repeal.
Deployable blend (anchor on open, tilt by model, weights fit 2022–23,
validated on 2024+2026): ~190 bets/yr, +4.95 pts CLV/bet (t = 9.6),
ROI +5.4% (t = 1.2 — positive expectation, slow to prove).
vs the closing line: Stage 2 closes ~45% of the Glicko-to-market log-loss
gap; incremental gain over the recalibrated close is +0.00075 pooled
(t = 1.68, n = 5,466) — real-looking, below significance, era-dependent.
Full analysis: RESULTS-v2.md. Rating-engine math: mma-glicko2-spec.md.
Betting system: BETTING-SYSTEM.md (KISS rule, validated 2012–2026, t=2.42).
Pipeline
pipeline/session.py UFC Stats HTTP session (SHA-256 proof-of-work solver)
cadence.py The weekly loop, automated (Thursday report, Sunday
settle); scheduled by .github/workflows/cadence.yml.
See site/README.md "Weekly cadence (automated)"
pipeline/scrape_v2.py A: event tables -> fights_v2.csv (per-fight KD/str/TD/sub)
B: fighter pages -> fighters_v2.csv (DOB/height/reach/stance)
C: fight details -> fight_details_v2.csv (attempts, control,
target/position splits; corner-order fixed by name key)
new: A + B for what is not on file yet, then C (weekly)
pipeline/scrape_bfo.py BestFightOdds line histories -> bfo_lines.csv
(sitemap -> event pages -> /api/ggd, base64+ROT47 decoded;
validated open = first paired tick unless corrected
within the hour (open_suspect), last tick = close;
--rebuild rewrites every row from the cached series)
pipeline/scrape_bfo_props.py BFO prop line histories -> bfo_props.csv (rounds
totals, goes the distance; method with --types method).
Same open rule as scrape_bfo; --join -> bfo_props_joined.csv
with fight_id, a/b side for fighter props, and the
card_settle in-play-safe clean close (from cache only)
pipeline/features.py Stage 1 Glicko replay + leakage-safe rolling features
-> features_v2.csv
pipeline/stage2.py Antisymmetric feature matrix (paired diffs, mirror aug)
pipeline/walkforward.py Walk-forward eval vs recalibrated closing market
pipeline/clv_eval.py CLV eval vs opening lines (model-only probs; see the
circularity warning in the docstring)
pipeline/card_report.py Per-event report: model %, fair line, open/close, CLV,
results. Usage: card_report.py <bfo-slug> <YYYY-MM-DD>
For an event dated today or later it also records the
current line of every bout in placeable_lines.csv.
Lists the rule-fired picks with their placement stake.
Bouts BFO still lists that are off the live card
(withdrawals) go in OFF_CARD, keyed by (event date,
matchup id), and are dropped before the table, the
signals and the placeable capture.
pipeline/prefight_rd.py Pre-fight RD of both fighters per bout (time-inflated,
public scale) -> prefight_rd.csv; placement-policy input
pipeline/signals.py THE signal path: classify(bout state + line) ->
Decision(projectable, reason, signal, stake). The frozen
rule constants, pre-filter, pick'em tie-break, debut
rule, opener validity and placement policy v1, in one
place with their freeze dates. card_report, card_settle,
betting_system and site/build.py all call it;
tests/test_signal_path.py guards that they agree.
pipeline/names.py BFO -> UFC Stats name resolution: aliases
(data/name_aliases.csv), suffix and hyphen rules, exact
or joined match only; make_resolver fails loud on a
spelling near-miss (NameJoinError), a fighter not on
file is a debut
pipeline/betting_system.py Joins the data, runs signals.classify over every
bout and settles (flip>=65% + gap>=100pts, flat 1u;
placement policy v1). Usage: betting_system.py [year|all]
[--write-ledger]. VOID rows for signals only a suspect
opener fires; a bout with a debutant has no row.
pipeline/polymarket.py Polymarket moneyline per bout (second price source,
never a signal source): PM column + capture columns
in card_report; `polymarket.py backfill 2026` writes
pm_history/ and pm_clv_2026.csv
Run order: scrape_v2.py A -> B -> C -> scrape_bfo.py 2010 2027 ->
features.py -> walkforward.py -> clv_eval.py -> prefight_rd.py.
python replay_check.py # merge gate: must pass (after any re-scrape, before committing data)
All scrapers are disk-cached (cache/, cache_bfo/, cache_pm/) and
resumable; re-runs cost zero network requests. Be polite: built-in delays, do
not remove.
Data
data/fights_v2.csv 8,962 bouts 1994-2026 (through 2026-10-24), chronological bout order,
missed-weight flags (206 incidents)
data/fighters_v2.csv 2,774 fighters
data/fight_details_v2.csv 8,941 bouts: attempts, control time, strike splits
data/features_v2.csv model-ready pre-fight features (leakage-safe)
data/market.csv vig-free closing probs (public odds dataset join)
data/bfo_lines.csv 8,010 bouts: opening + closing mean lines 2010-2026;
the open is the validated open, with open_suspect
and the raw first ticks (f1_open_raw, f2_open_raw)
data/bfo_joined.csv bfo_lines joined to fights_v2 by fight_id (a/b
orientation), with open_suspect and a/b_open_raw
data/bfo_props.csv 15,431 prop lines 2015-2026 (raw scrape): O/U 1.5/2.5/
3.5/4.5 rounds and goes the distance, open + close
cross-book mean, side 1 = Over/Yes. GTD missing on
<2% of bouts/yr 2015-21, 4-5% 2022-26, 13% in 2024
(BFO returns 500 for those histories)
data/bfo_props_joined.csv bfo_props joined to fights_v2 (fight_id, side) plus
p1/p2_close_clean; cancelled/re-booked bouts dropped
data/name_aliases.csv bfo_name -> ufcstats_name, hand-maintained (names.py);
an empty ufcstats_name marks a distinct person
data/model_only_preds.csv walk-forward model-only probs (no market input)
data/clv_eval_full.csv per-fight CLV records
data/walkforward.csv per-fight loss diffs vs recalibrated close
data/prefight_rd.csv pre-fight RD of both fighters per bout (time-inflated,
public scale) from prefight_rd.py; placement-policy input
data/placeable_lines.csv current BFO line per bout, appended by
card_report.py when a future event is reported
(append-only: one row per bout per run, skipped when
the lines are unchanged); the bet side's earliest
capture is the ledger's placeable_line. Since
2026-09-30 also pm_p_self (fighter1's Polymarket
price), pm_liquidity, pm_slug; empty when no market
data/pm_history/<slug>.json hourly Polymarket price history of both outcomes
of a resolved 2026 ledger bout (polymarket.py backfill)
data/pm_clv_2026.csv per 2026 ledger pick: Polymarket price at our capture
time and last pre-fight price, PM CLV next to BFO CLV
data/system_ledger.csv ledger of record. GENERATED by
`betting_system.py <year|all> --write-ledger`;
never hand-edit it, regenerate it. Columns:
event_date, event, fight_id, fighter, rule, live,
units, rd_self, rd_opp, placed, stake_u, pnl_placed,
open_line, placeable_line, clean_close, clv_pts,
result, method, round, pnl, pnl_at_open, reason,
pm_p_at_capture (reason set = VOID row, not counted)
data/legacy_favtier_ledger.csv
abandoned favorites-tier system (3u/2u/1u), ends
2026-03-28 plus the Paris FLIP row; kept for the record
Validation discipline
- Stage 2: train ≤2022, tune 2023–24, holdout 2025+ — 2025 remains sealed
(2026 was spent as a test window; documented in RESULTS-v2.md).
- Walk-forward: every prediction from a model trained strictly on prior fights.
- CLV: model-only probabilities (stacked probs contain the close — circular).
- Engine replay reproduces the frozen v1 ratings bit-for-bit (max Δ 0.05).
v1 engine
engine.py + best_params.json (tau=0.2, sigma0=0.06, S_udec=0.80,
S_sdec=0.55). Key finding: split decisions carry almost no skill signal.
