Polymarket 订单簿数据
每一次订单簿更新，每侧 25 档。

l2.parquet — one row per book update per outcome token
 
ts_recv_ns          int64    our receive time, ns since epoch (authoritative order)
ts_src_ms           int64    the venue's own timestamp, ms
condition_id        str      the market
token_id            str      the outcome (YES or NO leg) — read as a string
event_kind          str      "snapshot" (venue resent the whole book) | "update"
source_event_type   str      upstream message: book | price_change
msg_seq             int64    capture sequence number
bid_levels          int16    price levels in the WHOLE bid book (the file keeps 25)
ask_levels          int16    price levels in the whole ask book
best_bid, best_ask  float64  top of book, rebuilt from the levels
payload_best_bid    float64  top of book as the venue's message declared it
payload_best_ask    float64
bid_ask_reconciles  bool     rebuilt top of book == declared top of book
bid_px_1 … bid_px_25   float64  price per level, best first, NaN past the book
bid_sz_1 … bid_sz_25   float64  size per level, in shares (contracts)
ask_px_1 … ask_px_25   float64
ask_sz_1 … ask_sz_25   float64

# pm_depth.py — how deep was a Polymarket order book, level by level and near the mid?
# Runs on the free sample bundle: https://tickfoundry.com/samples
import numpy as np
import pandas as pd
 
l2 = pd.read_parquet("l2.parquet")                 # one full 25-level snapshot per book update
markets = pd.read_csv("reference/markets.csv")
 
# The YES leg of the day's busiest market, over the four hours its question covers.
busiest = l2.groupby("condition_id").size().idxmax()
m = markets.set_index("condition_id").loc[busiest]
book = l2[l2.token_id == str(m.yes_token_id)].sort_values("ts_recv_ns").reset_index(drop=True)
end = pd.Timestamp(m.end_time)
start = end - pd.Timedelta(hours=4)
 
# Sample the book once a second so every statistic is time-weighted:
# a burst of 50 updates inside one second counts as one second, not as 50.
grid = np.arange(start.value, end.value, 1_000_000_000)
at = np.searchsorted(book.ts_recv_ns.to_numpy(), grid, side="right") - 1
snap = book.iloc[at[at >= 0]]
snap = snap[snap.best_bid.notna() & snap.best_ask.notna()]      # drop one-sided seconds
 
px = {s: snap[[f"{s}_px_{i}" for i in range(1, 26)]].to_numpy() for s in ("bid", "ask")}
usd = {s: px[s] * snap[[f"{s}_sz_{i}" for i in range(1, 26)]].to_numpy() for s in ("bid", "ask")}
mid = ((snap.best_bid + snap.best_ask) / 2).to_numpy()[:, None]
spread = (snap.best_ask - snap.best_bid).to_numpy() * 100     # cents; never the rounded spread column
 
print(m.question, f"· {start:%Y-%m-%d %H:%M} to {end:%H:%M} UTC")
print(f"{len(book):,} book updates for this token on the day, "
      f"{((book.ts_recv_ns >= start.value) & (book.ts_recv_ns < end.value)).sum():,} inside the window; "
      f"{len(snap):,} two-sided seconds sampled\n")
 
print("spread (¢)  median", round(np.median(spread), 2),
      "| ≤1¢", f"{np.mean(spread <= 1.0001):.1%}",
      "| 1–2¢", f"{np.mean((spread > 1.0001) & (spread <= 2.0001)):.1%}",
      "| 2–5¢", f"{np.mean((spread > 2.0001) & (spread <= 5.0001)):.1%}",
      "| >5¢", f"{np.mean(spread > 5.0001):.1%}\n")
 
rows = []
for c in (1, 2, 5):                                         # USDC resting within c cents of the mid
    bid = np.nansum(np.where(px["bid"] >= mid - c / 100 - 1e-9, usd["bid"], 0), axis=1)
    ask = np.nansum(np.where(px["ask"] <= mid + c / 100 + 1e-9, usd["ask"], 0), axis=1)
    rows.append({"within": f"±{c}¢", "bid median $": np.median(bid), "bid p10 $": np.percentile(bid, 10),
                 "ask median $": np.median(ask), "ask p10 $": np.percentile(ask, 10)})
print(pd.DataFrame(rows).round(0).to_string(index=False), "\n")
 
rows = []
for lvl in (1, 2, 3, 5, 10, 25):                            # size at each level, and through it
    rows.append({"level": lvl,
                 "bid $ at level": np.nanmedian(usd["bid"][:, lvl - 1]),
                 "bid $ through": np.median(np.nansum(usd["bid"][:, :lvl], axis=1)),
                 "ask $ at level": np.nanmedian(usd["ask"][:, lvl - 1]),
                 "ask $ through": np.median(np.nansum(usd["ask"][:, :lvl], axis=1)),
                 "ask levels present": f"{np.mean(~np.isnan(px['ask'][:, lvl - 1])):.0%}"})
print(pd.DataFrame(rows).round(0).to_string(index=False), "\n")
 
print("price levels in the whole book (the file keeps the best 25 a side): median bid",
      int(np.median(snap.bid_levels)), "· ask", int(np.median(snap.ask_levels)))

免费账户可在 2026 年 7 月的任意日期中领取 5 个市场·日，schema 完全相同，pm_depth.py 无需改动即可在每一个上运行。无需绑卡。