Skip to main content

Unreleased changes

Merged to main since 0.1.5 and not yet released. Released versions are on the Changelog.

Unreleased​

Added​

  • CFB / NFL: the advanced box's team rows carry yards_per_dropback, sack-inclusive yards per dropback: (pass_yards + sack_yards) / passes, over the same dropbacks as yards_per_pass, which is unchanged and still counts a sack as 0 yards. It is the basis of SDV's yardsdropback percentiles, so Game on Paper's Binion box can rank the number it shows against them. The espn_{cfb,nfl}_adv_team (and CFB adv_team_gamelog) releases pick it up on their next rebuild. (#740)
  • CFB / NFL: the advanced box's turnover rows carry passes_defended, the opponent's ESPN box passes defended against the team's throws (CFB adds Int; ESPN's college PD excludes it), null where the box has none. (#737)
  • CFB CBS ids: the CBS scoreboard resolver reads FCS and postseason pages for the id map, a new offline _match_cbs_card joins a season's stored CBS cards to ESPN games (either home/away order, within three days), and the ESPN -> CBS team map gains five FCS schools (325 teams). (#815)
  • Cricket: cricket_win_probability selects the ODI fielding-restriction regime from the match date (cricket_match_state now carries match_date; cricket_model_constants.ODI_REGIMES / cricket_regime: pre-2005, 2005 powerplays, 2011 window, 2012 two new balls, 2015+) and applies a regime whole -- its own resource grid and REGIME_CONSTANTS par/sigma -- or falls back to the pooled 2002-2026 fit (a grid with fewer than 100 cells of 30+ states, or no fitted constants: the 13-month 2011 window). Output gains regime.
  • Crosswalk history harvest: private sportsdataverse._crosswalk_history captures Fox, Yahoo and CBS raw game listings and per-game box payloads into the *-alt-raw repos: resumable, atomic, season-checked, and paced by XWALK_RPS / XWALK_<PROVIDER>_RPS only; a 403 or 429 stops the run. (#805)
  • Czech Extraliga: 9 elh_* wrappers in sportsdataverse.hockey.elh (play-by-play, shots, events, scoreboard), plus elh_game_ids. (#826)
  • eSports hokejovyzapis: 41 wrappers over the DEL (12), ICE Hockey League (12), EIHL (9) and Metal Ligaen (8) feeds, in sportsdataverse.hockey.{del_,icehl,eihl,metal}. (#826)
  • Hockey: sportsdataverse.hockey.sihf wraps the keyless Swiss Ice Hockey Federation statistics API (data.sihf.ch), covering every SIHF league (National League, Swiss League, MyHockey League, the amateur, junior and women's leagues). There are eleven sihf_* wrappers. Four return season tables: sihf_results, sihf_standings, sihf_player_stats and sihf_goalie_stats. sihf_phases lists the phase_id values those four take for a league and season (the ids differ by league, season and table), with each phase's name and game count. Six read one game: sihf_pbp (goals, penalties, goaltender changes, revoked goals and shootout attempts), sihf_game_info, sihf_game_player_stats, sihf_game_goalie_stats, sihf_game_team_stats and sihf_game_lineups. SIHF answers an unknown season, league or phase with another one (Season=1890 returns the current season), and only the body's filter echo shows it. The runtime compares that echo with the request and raises NoDataError when they differ, so a wrong season is never returned under the requested label. sihf_standings counts every game by default (venue="All"; asked nothing, SIHF answers the away-games table), and a table body with fewer rows than it announces raises AssetFetchError.
  • HockeyAllsvenskan: 11 ha_* wrappers in sportsdataverse.hockey.ha (schedule, play-by-play, standings, skater/goalie/team stats, leaders, players). (#826)
  • KHL: 23 wrappers in sportsdataverse.hockey.khl: 14 khl_* over the khl.ru REST and 9 khl_webcaster_* over the webcaster API. (#826)
  • Ligue Magnus: 11 magnus_* wrappers in sportsdataverse.hockey.magnus (competitions, schedule, standings, game sheet, play-by-play, lineups, shootout). (#826)
  • Liiga: 23 liiga_* wrappers in sportsdataverse.hockey.liiga (schedule, box scores, shot map, standings, team and player stats). (#826)
  • MLB: load_mlb_schedule(seasons) reads the mlb_schedules release (statsapi schedule, every game type, 1988+).
  • MLB: load_mlb_statcast(seasons) reads the mlb_statcast release (Baseball Savant pitch-level data, one parquet per season, 2015+; spring training and exhibition pitches included). The on_1b / on_2b / on_3b runner ids are published as floats and cast to Int64 at the loader boundary.
  • National League: 17 nl_* wrappers in sportsdataverse.hockey.nl (games, box scores, shots with xG, standings, player and team stats). (#826)
  • NBA / WNBA / MBB / WBB crosswalks: {nba,wnba,mbb,wbb}_schedule_crosswalk(fox=True) fills fox_game_id (and the NBA/WNBA fox_home_team_id / fox_away_team_id) from Fox's per-date scoreboard, matched on US-Eastern date and teams in either home/away order (a late West-coast tip ESPN files at 23:59 ET matches Fox's 00:00 ET a day apart); nba_crosswalks_from_archive / wnba_crosswalks_from_archive(fox_games=True) fill them for any season Fox serves. yahoo=True on the WNBA / MBB / WBB team and schedule crosswalks fills yahoo_team_id (Yahoo's listing joined to ESPN by full name, short name, then nickname where the school names agree; unmatched Yahoo teams are logged) and yahoo_game_id = {lg}.g.{YYYYMMDD ET}{home Yahoo id}, checked against the ids Yahoo lists so a neutral-site game takes Yahoo's home side. Both legs are opt-in and fail-soft: any Fox or Yahoo failure logs a warning and leaves the columns null.
  • NBA / WNBA crosswalks: nba_crosswalks_from_archive / wnba_crosswalks_from_archive take fox=True: fox_team_id by team name for any season, and the Fox player columns for the current season only (Fox rosters are today's). A failed Fox fetch warns and leaves them null. (#786)
  • NBA / WNBA: nba_crosswalks_from_archive() / wnba_crosswalks_from_archive() build a past season's team, schedule and player crosswalks offline from the ESPN schedule and game-roster parquet plus the local stats box-score archive; the live builders only reach the current season.
  • NBA: load_nba_stats_draft(seasons) reads the nba_stats_draft release; seasons is the draft year. (#780)
  • NBA: loaders for the multi-file nba_stats releases: load_nba_stats_synergy, _hustle_player / _team, _matchups / _matchups_rollup and five _draft_combine_* tables. Each unions a season's files, stamps season_type / per_mode from the file name where rows lack them, and takes START-year seasons. (#780)
  • NBA: the team and schedule crosswalks now compute yahoo_team_id (nba.t.{espn id}) and yahoo_game_id (nba.g.{YYYYMMDD ET}{home id}) from 2014 on, where they were null.
  • NCAA: scrape.ncaa.espn_player_xwalk maps NCAA roster player ids to ESPN athlete ids per (season, ESPN team) for MBB, WBB and MFB, in four tiers that leave ties unmatched; the -raw repos run it as stage 07. (#813)
  • SHL: 27 shl_* wrappers in sportsdataverse.hockey.shl, plus shl_ssgt_uuid and shl_player_crosswalk. (#826)
  • SIHF: 22 more sihf_* table wrappers (team, skater and goalie goals, shots, special teams, penalties, shootouts, faceoffs, time on ice, attendance), 33 in the family. (#826)
  • VHL: 13 vhl_* wrappers in sportsdataverse.hockey.vhl (schedule, game report, play-by-play, standings, skater/goalie stats, players, rosters), plus vhl_tn. (#826)
  • WNBA crosswalks: wnba_crosswalks_from_archive(..., yahoo=True) fills yahoo_team_* and yahoo_game_id for a past season through the live builders' Yahoo code; a failed Yahoo fetch leaves the columns null with a warning. (#803)
  • _crosswalk_alt.match_season(known_teams=) takes fixed ESPN -> provider team pairs, and known_teams("cbs", "cfb") serves the CFB CBS resolver's static map. Date co-occurrence cannot separate teams that share every game date: an FCS team with one FBS opponent, or any team early in a season. With the map, CBS CFB games match 99% of ESPN's schedule in most seasons (2026 so far: 101 to 445 of 448), and matched players rise from about 7,500 to 10,000-12,300 a season.
  • _crosswalk_alt.with_alt_game_legs() / fill_game_legs() fill a basketball schedule crosswalk's Fox game and team ids (and any null Yahoo game id) from the parsed *-alt-data history, only where the game id is null. On the published NBA crosswalks it fills 1,260 of 1,379 games for 2002 and 1,326 of 1,405 for 2025; the rest are preseason and All-Star games, which Fox does not list.
  • nba_crosswalks_from_archive() / wnba_crosswalks_from_archive() take alt_root=: the *-alt-data checkout whose parsed Fox / Yahoo history fills the player crosswalk's fox_athlete_id / fox_player and yahoo_player_id / yahoo_player_name offline, only where they are null. The new private sportsdataverse._crosswalk_alt parses the stored Fox / Yahoo / CBS payloads into per-season games and game_rosters parquet (python -m sportsdataverse._crosswalk_alt) and matches them to ESPN: teams by (date, side) co-occurrence, games on (US-Eastern date, team pair), players inside each game and team, voted across the season.
  • ncaa_wbb_shot_locations takes season=, resolved through wbb_period_model (halves through 2014-15, quarters from 2015-16) instead of binding quarters unconditionally; the default stays quarters because stats.ncaa.org shot locations only exist from 2019. (sdv-internal-refs rules/wbb.yaml#wbb-2016-four-quarters)
  • parse_soccer_scoreboard (every espn_<league>_scoreboard) carries the event's season_type and season_slug, and both it and the summary header frame carry a derived tournament column ("apertura" / "clausura", null for single-tournament leagues) so Liga MX's two competitions per ESPN season id can be split. (sdv-internal-refs rules/ligamx.yaml#ligamx-1996-two-short-tournaments-per-season)
  • sportsdataverse.nba.nba_game_page: per-play in-game win probability with deltas and an overtime regime, the pregame prior from the spread, shot coordinates in the court frame, team four factors / ratings / pace, and player TS% / usage (NBA 48-minute and WNBA 40-minute regulation via league_id; only "00" and "10" are scored, anything else raises ValueError), for the Game on Paper NBA game page and the sdv-db nightly stage. Team turnovers repair ESPN's duplicated pre-2013 team turnover (team_turnovers == turnovers, total_turnovers == 2 * turnovers) and its null 2018 total_turnovers, and the win probability settles to 1.0 / 0.0 on the "End Game" / "End Period" row of a decided game from the 4th period on. A null score on a play is carried forward before scoring, and game_page_frames returns the rate frames with their documented columns (zero rows) before both team box rows exist.
  • NFL Fox event ids resolve through sportsdataverse.nfl.fox_pbp.game_id._resolve_fox_game_id, the Fox twin of the CBS week-scoreboard resolver (same arguments, light import), so sdv-db's nightly id map can fill fox_event_id; the football id map's team table now carries fox_team_id.

Changed​

  • CFB validation: the box-score parity harness grades penalties (accepted fouls, committing team), fumbles lost (offense and defense) and net passing (total minus rushing) the way ESPN counts them, and adds a player-level check and a per-game integrity screen. (#764)
  • CFB: the team summaries, percentiles, passing, receiving and advanced-box column descriptions match the republished definitions: run-stuff and opportunity rates over rushes, per-drive drive metrics, passer success over dropbacks, per-target receiving rates null for 2021-2024. (#741)
  • MBB/WBB: ncaa_espn_team_crosswalk(league=) reads the {mbb,wbb}_crosswalk sportsdataverse-data release (built weekly by sdv-reference-data), so a new season no longer waits for an sdv-py release; it warns and falls back to the bundled CSV when the release cannot be read. (#770)
  • MBB: load_mbb_player_crosswalk() documents kp_player_id, the KenPom player id hoopR-mbb-data now publishes after yahoo_player_name (18 columns), and kp_game_id on load_mbb_schedule_crosswalk() is described as filled. The live mbb_player_crosswalk() builder stays at 17 columns. (#787)
  • MLB: fit_zone_model (the umpire-zone and catcher-framing called-strike logistic) excludes 2026+ pitches with a UserWarning -- Savant reports their location at the middle of the plate against the ABS zone and the calls are challengeable -- and mlb_umpire_called_strike_prob / mlb_umpire_bias / mlb_catcher_framing carry an abs_era flag (per-umpire and per-catcher rows split by era).
  • Possession cache: nba_season_compile.PIPELINE_VERSION is 5, so per-game caches keyed by it (compile_nba_season, and WNBA builders that import _game_cache_key) recompile once. (#818)
  • Spring football: enrich_spring_football_pbp threads the league-season touchback_yardline and ep_point_values from get_sf_constants into enrich_nfl_pbp (new keyword-only league=, touchback_yardline=, ep_point_values=; NFL defaults unchanged), so XFL / UFL kickoffs are scored from their own touchback spot.
  • Tests: warnings are errors. A test asserts an expected warning with pytest.warns(..., match=...) or filters an incidental one by exact message; live tests skip on a timeout or upstream 429/5xx, the live job runs on Ubuntu, and the tests extra needs pytest >= 8.0. (#726)
  • load_cfb_teams_crosswalk / load_cfb_schedule_crosswalk accept 2013. The cfb_crosswalk release now carries 2013 (Yahoo and Fox legs), which the old 2014 floor hid.

Deprecated​

  • NBA / WNBA: load_nba_stats_{pbp,possessions,lineups}_v3 and the four load_wnba_stats_* shims (lineups, player_season_stats, standings, team_season_stats) now say they will be removed in 0.2.0; they named 0.1.0, which has shipped. A test fails if a removal version falls behind again. (#726)

Removed​

  • MBB / WBB: the NCAA <-> ESPN team-crosswalk builder tools/crosswalk/ is gone; sdv-reference-data (sdv_reference/crosswalks/) builds and publishes the {mbb,wbb}_crosswalk asset that ncaa_espn_team_crosswalk() reads first (#770). The bundled fallback copies stay. (#788)

Fixed​

  • CFB / NFL: the advanced box's expected_turnovers takes its pass term from ESPN's box passes defended, not from "broken up by" text whose coverage swings by season (2025 CFB weeks 1-8 vs 9-15: 0.68 vs 1.45 before, 1.26 vs 1.46 now); text stays the fallback where the box has no PD. (#737)
  • CFB play-by-play: a fumble recovered by a team written mixed-case ("recovered by AlbSt") or in parentheses ("recovered by Omar Jacobs (BGSU)") is charged to the right team; 2014-20 team fumbles lost now match ESPN's box in 97-98% of team-games, from 86-90%. (#767)
  • CFB play-by-play: kneel-downs ("Kneel down by TEAM ... (team loss of 2)", "takes a knee") and rushes no yardage reader matched (2004-13 "rush up the middle for 21 yards", yardless vendor rows) carry rushing yards; 2025 team rushing yards match ESPN's box in 64% of team-games, from 46%. (#768)
  • CFB play-by-play: pre-2014 extra-point rows ("Extra point by Dave Rayner (MSU) is good.", "Felipe Macias extra point GOOD.") set xp_attempt, xp_made and xp_kicker_player_name; all 46,762 had none. (#771)
  • CFB play-by-play: pre-2014 penalties ("HAWAII penalty 15 yard personal foul accepted") are charged to the team named in the text; 2006-13 team penalties match ESPN's box in 72-83% of team-games, from 21-25%. (#769)
  • CFB play-by-play: regex-fallback player names no longer carry the clause they ran into ("Tommylee Lewistackled b", "Logan Hall, clock:32", a spot read as a fumbler); garbled names fall below 0.3% in every role and era, from up to 21%. (#772)
  • CFB: cfb_rosters_crosswalk(providers=(..., "yahoo")) fills yahoo_athlete_id from Yahoo's team roster, matched by name + jersey, then by a unique name; it read the ~200-player stat-leaders feed and found 0-2 per team. New optional yahoo_team_id skips the Yahoo team lookup. (#762)
  • CFB: get_go_wp returns NaN for a 4th-down row with no yards_to_goal or distance, instead of a garbage go-for-it value from casting NaN to an integer; other plays are unchanged. (#726)
  • CFB: overtime win probability values the forced two-point try after a touchdown (the third overtime on before 2021, the second from 2021; cfb_wp_overtime.OT_TWO_POINT_FORCED_FROM) instead of a kicked extra point in every period; get_2pt_probs recommends the try there with xp_wp null.
  • CFB: scoring zero rows (QBR on a live game's opening drive, predict_from_card on an empty frame) no longer logs XGBoost's "Empty dataset" warning. (#726)
  • Crosswalks: {mbb,wbb,nba,wnba}_player_crosswalk(season=) for a past season reads that season's ESPN player box scores; ESPN's team-roster endpoint ignores ?season=, so past seasons shipped the current roster under their own label (MBB 2025, WBB 2014-2025). (#782)
  • ESPN rankings parsed to an empty frame. espn_{cfb,mbb,wbb,mch,wch}_rankings() ran the Site v2 payload through parse_items, which reads a Core v2 items list; the polls sit in a top-level rankings list, so every league returned zero rows. A dedicated parse_rankings now gives one row per ranked or vote-receiving team per poll (live 2026-10-08: cfb 208 rows over 5 polls, mbb 77, wbb 75, mch 35, wch 20), with poll_id, season, season_type, week (the week within the season type, which ESPN's Core v2 rankings URL uses), rank (null on vote-receiving rows), previous_rank, points, first_place_votes and the team fields; team_id is a string. The college baseball and softball wrappers route through it too, but ESPN answers their rankings URL with a 404. (#732)
  • Hockey: sihf_phases gives a null phase_name when a phase's name is missing, blank or not a string, like the other SIHF parsers; it gave "" for a missing or blank name and the stringified value ("['x']") for a list or object.
  • HockeyTech derived coordinates put home-team events off the rink. add_coord_transforms (every <league>_pbp and pwhl_pbp) flipped home events with a 0-200 x 0-85 mirror applied to centre-origin feet, so home x_coord_right ran 191 to 290 ft and y_coord_right 49 to 125 ft (PWHL game 42), and y_coord_fixed applied the y transform twice. The feed puts the home team's attack toward x = 0 in every period (320 PWHL games; sdv-internal-refs #50), and the columns are now rotations of the feet frame (#51): x_coord_fixed / y_coord_fixed = (-x, -y), home team shooting right; x_coord_right / y_coord_right = home (-x, -y), visitor (x, y); x_coord_vertical / y_coord_vertical = (-y_right, x_right). The right and vertical columns are null when an event's side is unknown (no team, as on faceoffs, or a null or empty home team id) instead of defaulting to the visitor's. shot_distance, shot_angle and xG are unchanged. The published PWHL play-by-play releases keep the old values until they are rebuilt.
  • HockeyTech empty-net goals from a team's own half were measured to the wrong net. add_shot_distance_angle measures to the nearer net (89 - |x|), which per-event validation showed is right for own-half events with a goalie in net: they are near-net events whose coordinates the feed mirrored (own-half quality shots match ordinary ones at 10 / 24 / 35 ft). An own-half empty-net goal is a genuine long shot (27 in 320 PWHL games; median 126 ft, measured 59 ft). Empty-net goals (empty_net "1") now measure to the net their team attacks; every other event is unchanged (sdv-internal-refs #52). shot_distance, shot_angle, scoring_chance and the xG features of those goals change.
  • HockeyTech schedules returned other seasons' games. <league>_schedule(season=...) for the 19 HockeyTech league families, and pwhl_schedule, read modulekit/scorebar. That view ignores the season_id it is sent, returns games oldest-first and stops at its 10,000-row limit, so ahl_schedule(season=2026) returned 10,000 AHL games from 1995 to 2012 and none from 2025-26; OHL and ECHL failed the same way, and PWHL returned every season's games. The schedule functions now read modulekit/schedule&season_id=<id>, which the feed scopes to one season (live 2026-10-08: AHL 2025-26 = 1,152 games, OHL = 682, PWHL = 120, each in that season alone), and parse_schedule(..., season_id=) drops any row from another season (#727). Behaviour changes:
    • With neither season nor season_id, the schedule is the newest regular season, as in the other season-aware functions. It used to be the feed's whole window.
    • game_status now reads the schedule view's strings: Final OT where scorebar said Final 1st OT, and the start time (7:00 pm EST) for an unplayed game.
    • A season's playoffs and preseason are separate HockeyTech season ids; pass season_id for them.
    • An empty reply, or a season filter that keeps nothing, is zero rows with the 12 columns.
  • MLB docs: the pitcher column in the load_mlb_xera, load_mlb_stuff_plus, load_mlb_command_plus, mlb_statcast_search (+ _minors / _wbc), as_of_split and bip_trajectory_features returns tables now reads "MLBAM player id of the pitcher" instead of the inherited "Whether the position is a pitcher". On the Statcast-search, as_of_split and bip_trajectory_features pages, home_score / away_score are the score before the pitch (not after the play) and player_name follows player_type (the batter by default). On espn_mlb_schedule, home_score / away_score are now described as the ESPN scoreboard score rather than a per-play run total.
  • MLB: pitch_features standardizes velocity, spin, movement and release within pitcher AND season (game_year, else the year of game_date), so PITCHf/x speeds (2008-16) and Statcast out-of-hand speeds (2017+, about +0.5-1 mph) never share one mean.
  • NBA / WNBA crosswalks: the archive team and player crosswalks no longer carry All-Star and other exhibition sides (e.g. WNBA 2026 TEAM SPOON / TEAM COOP); their games stay in the schedule crosswalk. (#786)
  • NBA: logistic_fit_irls no longer overflows on separable data, and nba_rapm fits a single possession without a divide-by-zero warning. (#726)
  • NFL: get_go_wp on a full nflverse frame no longer raises pandas PerformanceWarnings; output is unchanged. (#726)
  • NFL: overtime wp / vegas_wp now follow the season's overtime rule format (model_vars.OVERTIME_FORMATS): sudden death, nflfastR's One_FG form with a per-game Drive_Diff, or a both-teams-possess form (postseason 2022+, regular season 2025+) under which an opening-drive touchdown no longer ends the game; spring-football overtime gets null.
  • NFL: the kickoff-touchback spot is now the rule spot per season (model_vars.touchback_yardline: 80 / 75 / 70 for 2024 / 65 from 2025) in the EP feature substitution, the overtime re-score and the ESPN yds_kickoff_return; nflfastR still scores 2024+ kickoffs at 75, so kickoff-row ep diverges from nflverse there by design.
  • NHL: load_nhl_game_rosters() accepts published roster seasons from 2010 instead of rejecting seasons before 2024.
  • NHL: twelve release loaders rejected seasons their releases publish: load_nhl_game_info(), load_nhl_goalie_boxscores(), load_nhl_linescore(), load_nhl_penalties(), load_nhl_scoring(), load_nhl_scratches(), load_nhl_skater_boxscores() and load_nhl_three_stars() started at 2024, and load_nhl_officials(), load_nhl_shifts(), load_nhl_shootout() and load_nhl_shots_by_period() at 2025. Every NHL release publishes from 2010 (the 2009-10 season), and all of them now accept it. load_nhl_pbp_lite() reads the release's current play_by_play_lite_<season> files, now published for 2010-2027. The legacy play_by_play_<season>_lite files stopped at 2026, so the current season was missing. The current files hold every legacy column, plus game_date.
  • sdv-docs index: stdlib and third-party objects a module re-exports (typing.Any, xgboost.Booster, tqdm, functools.lru_cache, 14 in all) are no longer indexed as sdv-py functions. SDV partials and type aliases whose __module__ is functools or typing stay indexed.
  • sdv-docs: the server opens an index whose path contains ?, # or %; such a path used to end the SQLite URI early and silently open a different, empty file. A corrupt or foreign SDV_DOCS_DB now gets an index unreadable: … answer instead of a raised DatabaseError.
  • sdv-docs: when the mcp import fails, sdv-docs prints the import error after the install hint, so an installed but incompatible mcp is visible.
  • WNBA / Summer League lineups: rotation lineups are timed on the league's own periods (WNBA 10-minute quarters, halves before 2006; Summer League 10-minute quarters), not NBA 12-minute ones; a malformed rotation returns an empty frame and every engine falls back to pbp lineups. (#818)
  • _crosswalk_alt CBS parse skips a box score whose playerStats is a list: CBS answers a few college football game ids with a basketball box (CFB 2017 and 2021), which used to stop the whole season's parse.
  • cfb_schedule_crosswalk builds the same rows from the same upstream data. A Yahoo scoreboard week that answered HTTP 200 with a cut-off JSON body was skipped, so its Yahoo ids and Yahoo-only games came and went between nightly builds (1,565 vs 1,608 rows for 2026). _download_json (and the Yahoo Shangrila getter) now re-fetches a truncated JSON body up to twice, and a full-season build raises CrosswalkSourceError on any failed ESPN calendar, ESPN week or Yahoo week instead of skipping it.
  • ncaa_wbb_game_pbp / ncaa_wbb_play_by_play bind the WBB period model for the game's era (halves through 2014-15, quarters from 2015-16) instead of quarters unconditionally; a pre-2016 game no longer parses to zero rows. New season= argument pins the era; without it infer_wbb_period_model reads it from the page's first-period clock. (sdv-internal-refs rules/wbb.yaml#wbb-2016-four-quarters)
  • parse_soccer_summary (so espn_<league>_summary for every ESPN soccer league) warns (UserWarning, naming the event id) when a payload carries a match header but key_events, commentary and lineups are all empty, instead of returning the zero-row frames silently. ESPN lists usa.nwsl 2014-2018 but ships no detail for those seasons. (sdv-internal-refs rules/nwsl.yaml#nwsl-2013-espn-season-coverage-start)
  • CFB pbp: fumble_lost / fumble_recovered read the recovering team from the text (a punt returner recovering his own fumble is no loss; a muffed punt the kicking team recovers is one), kickoff touchbacks come out to the 25 from 2012 (they switched only after 2013, in yds_kickoff_return and the EP touchback yardline), and yds_fumble_return reads the 2004-06 "returned for a N yard touchdown".
  • CFB pbp: pre-2014 onside recoveries (Kickoff Return (Defense)) are typed Kickoff before the kickoff possession flip, so they carry kickoff_play / kickoff_onside and the right kicking / receiving team; and td_play follows a touchdown type that 2014+ scoring-summary text never spells out ("12 Yd Interception Return (X Kick)").
  • Era-dependent behaviour catalogued in sdv-internal-refs rules/ is now honoured at the code sites it flagged: ESPN soccer summaries null the (0.0, 0.0) "no location" coordinate pairs and strip trailing spaces from names (rules/epl.yaml); cricket_match_state reads a DLS-reduced innings' revised over allocation from the score string instead of the format's full ball budget (rules/cricket.yaml#cricket-1999-duckworth-lewis); mlb_statcast_search clamps a pre-2008-03-25 window to the tracking floor with a warning and get_baselines warns when it falls back to the nearest committed season (rules/mlb.yaml); get_court(league_id, season=) returns the three-point geometry in force that season (NBA 22-ft arc 1994-97, WNBA 19'9" / 20'6.25" / FIBA eras); get_sf_constants(league, season=) resolves the UFL 2024 / 2025 / 2026, XFL and CFL rulesets; HockeyTech ot_period_length carries the sourced regular-season OT length for PWHL (300), ECHL and USHL (420) and BCHL (600).
  • NFL play-by-play opportunity_run now marks a carry that gained 4 or more yards; it was inverted (<= 4), so short runs counted as opportunities and the box's rushing_opportunity_rate ran about 0.62 instead of about 0.46. opportunity_run, highlight_run and adj_rush_yardage now gate on rush, as the CFB processor does, so rushing touchdowns and fumbled rushes are no longer dropped.
  • The archive crosswalks' team pairing (date and home/away co-occurrence) leaves a team unpaired when two candidates tie, instead of taking the lower id. A full season never ties; a slice of a few dates could pair the wrong franchises.
  • The fourth-down surface's PAT make probability (get_2pt_wp / get_2pt_probs, conv_1pt) scores the FG model at nflfastR's pre-2015 spot (yardline_100 = 3, the kick snapped from the 2) for seasons before 2015 instead of the modern 15 for every season; new PAT_YARDLINE_PRE_2015 / PAT_YARDLINE_POST_2015 / PAT_FIRST_SEASON_FROM_15 constants in nfl.model_vars. (sdv-internal-refs rules/nfl.yaml#nfl-2015-pat-snap-from-15)

Data​

  • Cricket: the win-probability artifacts are refit on Cricsheet male internationals 2002-2026 (3,581 T20I + 2,589 ODI): the surface is regime-keyed with per-cell support, the calibrator is fitted on the regime-aware raw probabilities, pooled constants move to T20 149.5/42/33 and ODI 249.3/76/59, and both holdout fixtures carry match_date. Holdout Brier 0.1898 -> 0.1882, max per-decile gap 0.0326 -> 0.0369 (gate 0.05).
  • NCAA 2026-27 MBB (365) and WBB (363) team ids added to the bundled ncaa_teamids_{mbb,wbb}.csv, captured from stats.ncaa.org's Division I team list, and their 2026-27 rows to the bundled NCAA ↔ ESPN team crosswalks (every team matched; West Florida joins in the ASUN, New Orleans is now LSU New Orleans, Saint Francis left Division I).