Datasets:

---

vgregoire

/

polymarket-users

like 10

Thanks for this dataset. While validating the user-level features (v1.2, commit 81cd110c) I found that three columns in user_features are the same column: frac_longshot, frac_sureshot, and frac_extreme_price are byte-identical, and all three equal the share of a user's fills at a price in either tail (price < 0.10 OR price > 0.90). The directional longshot (low-price) and sureshot (high-price) features appear to have collapsed into the two-tail extreme-price measure.

Minimal reproduction (a few seconds, reads only user_features):

import pandas as pd

cols = ["frac_longshot", "frac_sureshot", "frac_extreme_price"]
df = pd.read_parquet(
"hf://datasets/vgregoire/polymarket-users/user_features.parquet",
columns=cols,
)
print(len(df)) # 2,480,104
print((df.frac_longshot != df.frac_sureshot).sum()) # 0
print((df.frac_longshot != df.frac_extreme_price).sum()) # 0
print(df.nunique(axis=1).value_counts()) # every row: 1 distinct value
Deeper check against the trade tape. Reconstructing each user's share of fills at price < 0.10 OR price > 0.90 from the full trades tape (588,287,500 fills, counting both maker and taker sides) and comparing to the shipped frac_longshot, over all 2,480,104 users:

correlation = 1.000000, mean absolute deviation = 0.0, and the 99th-percentile absolute deviation = 0.0 (matches for every user);
other cutoffs don't match: 0.05/0.95 -> corr 0.933, 0.20/0.80 -> corr 0.917;
a correct one-tail longshot share (price < 0.10 only) -> corr 0.806, MAD 0.181.
So the operative cutoff is 0.10/0.90, two-tailed and undirected. The two tails are genuinely distinct at the fill level (19.4% of fills at price < 0.10 vs 10.2% at price > 0.90), so this isn't a coincidence. frac_midrange does not appear affected.

Likely cause (guessing at the pipeline here): all three columns seem to have been assigned the same is_extreme = (price < 0.10) | (price > 0.90) computation, overwriting the intended mean(price < 0.10) and mean(price > 0.90).

Happy to share the full reproduction script. Thanks again for the dataset, is a fix planned so I can pin the corrected version?