DATA_DICTIONARY.md

Data Dictionary — Ringside Wrestling Archive

Column-by-column documentation for every table in the dataset. Types reflect the underlying Postgres schema; parquet preserves them faithfully (CSV exports drop type info — see notes below each table).

Conventions:

• All id columns are unsigned integers, monotonically increasing.
• All timestamp columns are UTC.
• "Nullable" = Y means the column can be NULL/empty for some rows.
• created_at / updated_at columns track ETL provenance, not real-world events.

---

1. promotions (6 rows)

The major North American promotions covered.

Column	Type	Nullable	Description
id	int	N	Primary key
name	text	N	Full promotion name (e.g., "World Wrestling Entertainment")
abbreviation	text	N	Common short form (WWE, AEW, WCW, ECW, NXT, TNA)
founded	date	Y	Founding date
defunct	date	Y	Closure date (null if active)
parent_org	text	Y	Parent corporation if applicable (WWE → TKO, NXT → WWE)
created_at	timestamp	N	When this row was first inserted
updated_at	timestamp	N	Last modified by ETL

Notes: WCW (1988-2001) and ECW (1992-2001) are historical cohorts — their match counts won't grow.

---

2. wrestlers (12,814 rows)

Identity table. One row per unique wrestler, regardless of name changes.

Column	Type	Nullable	Description
id	int	N	Primary key
ring_name	text	N	Canonical ring name (most-used name; alternates in wrestler_aliases)
real_name	text	Y	Birth/legal name where known
gender	text	Y	M, F, or Other. Null when undetermined from sources
birth_date	date	Y	Date of birth
debut_date	date	Y	First documented match in any promotion
status	text	Y	active, retired, deceased, unknown
primary_promotion_id	int	Y	FK → promotions.id (most-associated promotion)
brand	text	Y	Sub-roster (e.g., Raw, SmackDown, NXT) for current wrestlers
billed_from	text	Y	Storyline hometown ("billed from Cleveland, Ohio")
image_url	text	Y	Cagematch CDN URL (may rot — re-host before depending)
created_at	timestamp	N	Row insertion
updated_at	timestamp	N	Last update

Notes: Use wrestler_aliases to find rows for wrestlers who have used multiple ring names (Dwayne Johnson → Rocky Maivia / The Rock).

---

3. wrestler_aliases (13,230 rows)

Alternate ring names and their active periods.

Column	Type	Nullable	Description
id	int	N	Primary key
wrestler_id	int	N	FK → wrestlers.id
alias	text	N	The alternate name
promotion_id	int	Y	Promotion where this alias was used (null if cross-promotion)
active_from	date	Y	First documented use of the alias
active_to	date	Y	Last documented use (null if still in use)
created_at	timestamp	N	Row insertion

Notes: Useful for entity resolution when joining with external sources that use different name forms.

---

4. events (35,064 rows)

One row per show — TV taping, PPV, house show, indie event.

Column	Type	Nullable	Description
id	int	N	Primary key
name	text	N	Event name (e.g., "WrestleMania 40 Night 1")
promotion_id	int	N	FK → promotions.id
date	date	N	Event date
venue	text	Y	Arena/venue name
city	text	Y	City of event
state	text	Y	State/province (US/Canada only, mostly)
country	text	Y	Country code or name
event_type	text	Y	ppv, weekly_tv, house_show, nxt_show, indie, etc.
cagematch_id	int	Y	Cagematch.net's event ID — useful for re-fetching source HTML
created_at	timestamp	N	Row insertion
updated_at	timestamp	N	Last update

Notes: event_type is most reliable for WWE/AEW; territory-era events may default to weekly_tv or be null.

---

5. matches (482,166 rows)

Match-level metadata. Joins to participants via id.

Column	Type	Nullable	Description
id	int	N	Primary key
event_id	int	N	FK → events.id
match_order	int	Y	Position on the card (1 = opener, higher = later)
match_type	text	Y	singles, tag_team, triple_threat, fatal_four_way, royal_rumble, ladder, cage, etc.
stipulation	text	Y	Special rules (No DQ, Hell in a Cell, 2 out of 3 falls)
duration_seconds	int	Y	Match length where recorded
title_match	bool	Y	True if a championship is on the line
rating	float	Y	Cagematch user crowd rating, 0.00 to 10.00. Real human signal — not scripted.
cagematch_id	int	Y	Cagematch.net's match ID
created_at	timestamp	N	Row insertion
updated_at	timestamp	N	Last update

Notes:

• rating is the only column that's not a writer's decision — it's how viewers actually felt about the match. Useful as a target variable for regression problems.
• ~20% of rows have a non-null rating; coverage is best for PPV / weekly-TV singles matches and worst for territory-era house shows.

---

6. match_participants (731,133 rows)

The fact table. One row per (match, wrestler). result is the label for outcome prediction.

Column	Type	Nullable	Description
id	int	N	Primary key
match_id	int	N	FK → matches.id
wrestler_id	int	N	FK → wrestlers.id
team_number	int	Y	Team grouping for tag/multi-person matches; same number = same side
result	text	N	win, loss, draw, dq (disqualification), no_contest, countout
entry_order	int	Y	Entry position for Royal Rumble / battle royal style matches
elimination_order	int	Y	Order eliminated in multi-person matches (null if not eliminated)
created_at	timestamp	N	Row insertion

Notes:

• For singles matches there are exactly 2 rows; for tag/multi there can be many.
• The label distribution: ~46% win, ~46% loss, ~3% draw/dq/no_contest/countout combined. Most ML pipelines filter to win/loss only.
• Kayfabe warning: result records the booked outcome, not athletic ability.

---

7. titles (121 rows)

Championship belts.

Column	Type	Nullable	Description
id	int	N	Primary key
name	text	N	Title name (e.g., "WWE Championship")
promotion_id	int	N	FK → promotions.id
established	date	Y	First-awarded date
retired	date	Y	Date retired/unified (null if active)
active	bool	N	Whether the title is currently defended
created_at	timestamp	N	Row insertion
updated_at	timestamp	N	Last update

---

8. title_reigns (1,753 rows)

Reign-level history of championship holders.

Column	Type	Nullable	Description
id	int	N	Primary key
title_id	int	N	FK → titles.id
wrestler_id	int	N	FK → wrestlers.id
won_date	date	N	Date title was won
lost_date	date	Y	Date lost (null if current reign)
defenses	int	Y	Number of recorded defenses during the reign
won_at_event_id	int	Y	FK → events.id (event where reign began)
lost_at_event_id	int	Y	FK → events.id (event where reign ended)
created_at	timestamp	N	Row insertion
updated_at	timestamp	N	Last update

Notes: defenses is conservatively counted — only matches explicitly tagged as title defenses in source data.

---

9. alignment_turns (631 rows)

Face/heel/tweener transitions per wrestler.

Column	Type	Nullable	Description
id	int	N	Primary key
wrestler_id	int	N	FK → wrestlers.id
from_alignment	text	Y	Previous alignment (face, heel, tweener, null)
to_alignment	text	N	New alignment after the turn
turn_date	date	N	Approximate date of the turn
event_id	int	Y	FK → events.id (event where turn happened, if pinpointable)
description	text	Y	Storyline context ("turned heel after attacking former tag partner")
source	text	Y	cagematch, manual, derived
created_at	timestamp	N	Row insertion

Notes: Coverage is sparse (~631 turns vs. 12.8K wrestlers); only well-documented turns are recorded.

---

10. match_view.parquet (denormalized, 731K rows)

Pre-joined ML-ready table. No SQL joins required to use this for outcome prediction.

Column	Type	Nullable	Description
match_id	int	N	FK → matches.id
wrestler_id	int	N	FK → wrestlers.id
ring_name	text	N	Wrestler's canonical ring name
event_id	int	N	FK → events.id
event_date	date	N	Date of the match
year	int	N	Year extracted from event_date
event_type	text	Y	PPV/TV/etc.
promotion_id	int	N	FK → promotions.id
promotion_abbr	text	N	WWE, AEW, etc.
match_type	text	Y	Match format
stipulation	text	Y	Special rules
title_match	bool	Y	Title on the line
duration_seconds	int	Y	Match length
rating	float	Y	Cagematch crowd rating
team_number	int	Y	Team grouping
entry_order	int	Y	Royal Rumble entry
elimination_order	int	Y	Multi-person elimination order
result	text	N	The label — win/loss/draw/dq/no_contest/countout
n_participants	int	N	Total wrestlers in this match
n_teams	int	N	Distinct team_numbers in this match
is_singles	bool	N	True if n_participants == 2 and n_teams == 2

When to use: Prefer match_view for outcome modeling; prefer the source tables when you need full schema fidelity (aliases, alignments, title context).

---

11. feature_matrix.parquet (~480K rows, model-reproducible)

The 35-feature ML-ready matrix used by the trained xgboost.joblib model. Lets users reproduce model predictions without rebuilding the feature pipeline.

Identifier columns:

Column	Type	Description
match_id	int	FK → matches.id
wrestler_id	int	FK → wrestlers.id
event_date	date	Match date — use for temporal splits
is_win	int	Binary label (1 = win, 0 = loss)

The 35 features (grouped by family):

Win momentum (5):

• win_rate_30d, win_rate_90d, win_rate_365d — rolling win rates
• current_win_streak, current_loss_streak

Event context (4):

• is_ppv, is_title_match, card_position (1=opener), event_tier

Match type (9):

• match_type_win_rate — wrestler's win rate in this match type
• is_singles, is_tag_team, is_triple_threat, is_fatal_four_way
• is_ladder, is_cage, is_hell_in_a_cell, is_royal_rumble

Title proximity (3):

• is_champion, num_defenses, days_since_title_match

Career phase (3):

• years_active, matches_last_90d, days_since_last_match

Promotion (1):

• promotion_win_rate

Head-to-head (2):

• h2h_win_rate, h2h_matches

Alignment (6):

• alignment (categorical: face/heel/tweener)
• is_face, is_heel
• days_since_turn, turns_12m, face_heel_matchup

Match quality (1):

• avg_match_rating — wrestler's career-average Cagematch rating

Card position momentum (1):

• card_position_momentum — trend in main-event vs. opener placement

Important: All features are computed at pre-match time — no future data leakage. Computed by ml/features.py against the same Postgres schema used during training.

---

CSV vs. Parquet — what changes

CSV mirrors of all the above tables are provided for portability. Differences to be aware of:

Concern	Parquet	CSV
Type preservation	Yes (date, int, bool, float, text)	No — everything is a string
Nulls	Native NULL	Empty string
Booleans	True/False	True/False strings
Date parsing	Already datetime64[ns]	You must pd.to_datetime(...)
Compression	snappy (in-file)	None (Kaggle adds zip on upload)
Size on disk	5–10× smaller	Larger but human-readable

Recommendation: Use parquet for analysis; use CSV only if your environment lacks pyarrow/fastparquet.

---

Provenance

Sources:

• Cagematch.net (public HTML scrape, non-commercial use): bulk of post-1990 data
• alexdiresta/all-wwe-and-wwf-matches Kaggle dataset (profightdb dump): cross-validation + pre-1990 coverage

Normalization performed by the ETL pipeline at github.com/tedrubin80/wrastlingfirst.