# Schema

## depth/kalshi_depth_<YYYY-MM-DD>.parquet
One row per (snapshot, ticker, side, level). Snapshot cadence for this vintage, median snapshot interval measured at build time: 30.5 s before 2026-07-12; 43.0 s from 2026-07-12 onward (CAPTURE_STATS.json depth_cadence_s_measured). Every
output file has the same physical schema and column order shown below.
| column | type | meaning |
|---|---|---|
| timestamp | double | snapshot time, epoch seconds UTC |
| ticker | string | Kalshi market ticker (family = prefix, e.g. KXMLBGAME) |
| sport | string | capture label |
| market_type | string | GAME / MATCH / FIGHT / SPREAD / TOTAL / F5TOTAL / GSPREAD |
| line | double | line value for spread/total variants; null for moneyline and legacy rows captured before this field existed (the column is still physically present) |
| side | string | 'bid' or 'ask' |
| level | int16 | book level, 0 = best; **max 19 (20 levels/side cap)** |
| price_c | int16 | price in cents (1-99) |
| size | double | contracts at level. **Historical precision limit:** source depth captured through 2026-07-19 had already integer-truncated Kalshi fixed-point sizes; those fractional components are irrecoverable even though rebuilt files use float64. Captures after the recorder fix preserve fractions. |

## trades/kalshi_trades_<YYYY-MM-DD>.parquet
One row per unique executed print (deduped by trade_id), filed under the
UTC date of `created_time`.
| column | type | meaning |
|---|---|---|
| created_time | string | EXCHANGE trade time (UTC ISO-8601) — use this for analysis |
| trade_id | string | Kalshi unique print id |
| ticker | string | market ticker |
| price / no_price | int64 | YES / NO price in cents |
| count | double | contracts traded; rebuilt from the raw JSON capture, so fractional fixed-point quantities are genuinely preserved (do not cast to integer) |
| taker_side / taker_book_side / taker_outcome_side | string | aggressor labels |
| is_block_trade | bool | block-trade flag |
| game_id / game_state / score_diff | string | game context AT RETRIEVAL time, not event time ('' when absent; join to a score timeline at created_time for event-time analysis; no current lag distribution is asserted) |
| capture_ts | double | when OUR recorder observed the print (epoch sec) |

Joining prints to depth: match on ticker, align created_time to the
enclosing snapshot timestamps. Snapshot cadence bounds the alignment
precision — use the measured cadence for the regime you are joining in
(depth_cadence_s_measured in CAPTURE_STATS.json for the full tape, or in
manifest.json for a per-sport slice), not a nominal figure.