Quant data discrepancies
Hello! Thank you for your amazing and exhaustive dataset, it really helped me! However, I've found several schema discrepancies from the description and actual dataset.
Firstly, it seems that in "quant.parquet" token_id is changed to asset_id. The same can be said about maker_id and taker_id. There may be more of such cases that I missed. Although minor, I find it worth mentioning for the ease of use and planning.
Secondly, it would seem that "quant.parquet" contains several token_ids for a single market, containing both "YES" and "NO" token. I do understand that it holds significant practical value, though I find it rather confusing coupled with unified price and "normalized to token 1" mindset as it can lead to double counting in some cases and also requires a "cheatsheet" to remember which side the token is.
In any case, your dataset is still amazing and contains a lot of valuable data for all usecases, for which I am very grateful to you!
Thank you for your suggestions! The previous version did have some issues. We've released an updated version with the latest data through 2026-03-04.
Regarding quant.parquet: this is a derived dataset we built for our own quantitative research. It unifies all trades to the YES (token1) perspective — for token2 trades, the price is converted to 1 - price and the buy/sell direction is swapped. This is intentional by design, and there is no data duplication issue.
Regarding the field name discrepancies: I apologize for the confusion. Since quant.parquet was originally built for our internal use, some column names may have changed during processing. I've now added complete field-by-field schema documentation for all files in the README.
For general use, we recommend using trades.parquet directly — it preserves all original trade semantics.
Thanks again for the feedback! If you have any questions or ideas about Polymarket trading strategies, feel free to reach out — always happy to discuss.
Hi! Thanks for the dataset, very usefull and extremely exhaustive. I was wondering whether you will upload a new version of users.parquet, containing data also for January and February.
Thank you again for the amazing work.
Hi! Thanks for the dataset, very usefull and extremely exhaustive. I was wondering whether you will upload a new version of users.parquet, containing data also for January and February.
Thank you again for the amazing work.
Hi! Thank you for the kind words — really glad the dataset has been useful!
Regarding an updated users.parquet — I won't be publishing one. I do have a live pipeline continuously ingesting the latest Polymarket data, but given the significant growth in trade volume recently, processing and storing user-level aggregations has become quite heavy. More importantly, this data hasn't been directly relevant to my own strategy research, so I haven't been maintaining it.
If you need user-level activity data, trades.parquet might be a reasonable alternative — you can derive similar information by separating the taker and maker sides with some basic cleaning.
That said, if you're doing research specifically around user behavior on Polymarket, I'd be happy to hear more about what you're working on — there might be room for collaboration or exchange.
Thanks again for reaching out!
Thanks, I will make sure to let you know if I find something intresting in the dataset regarding users behavior. I had also a question regading the users.parquet. I might have found some duplicates, like trades that are exactly the same in everything (even in being taker or maker), is there a specific reason why this might happen? Thanks again.
Thanks, I will make sure to let you know if I find something intresting in the dataset regarding users behavior. I had also a question regading the users.parquet. I might have found some duplicates, like trades that are exactly the same in everything (even in being taker or maker), is there a specific reason why this might happen? Thanks again.
Thanks for catching this! You're right — we confirmed there are about 2.47% exact duplicate rows (5.58M out of 225.86M) in users.parquet.
This happened because during the early stages of data collection, the API endpoint was unstable and we had to restart the scraping process multiple times, which led to some overlapping data being recorded twice.
For now, you can simply deduplicate by dropping rows that are identical across all columns (a straightforward drop_duplicates() will do). We'll also clean this up in a future release.
Thanks again for flagging this — really appreciate it!
Hi! Thanks a bunch on this dataset! It's been really helpful. However we noticed some completeness issues comparing to other data sources, see the issue on github: https://github.com/SII-WANGZJ/Polymarket_data/issues/1
Hi! Thanks a bunch on this dataset! It's been really helpful. However we noticed some completeness issues comparing to other data sources, see the issue on github: https://github.com/SII-WANGZJ/Polymarket_data/issues/1
Thanks for the heads up! Really impressed by the thoroughness of your audit in #1 — working on the backfill now. Will update there.
Hello, thanks again for the dataset. I have a question about the dataset's functioning. Is it possible to compute the total profit by each user for each market? Can I capture the full wallet position, or can I only see the part of the wallet that is traded among users? Thanks again
Hello, thanks again for the dataset. I have a question about the dataset's functioning. Is it possible to compute the total profit by each user for each market? Can I capture the full wallet position, or can I only see the part of the wallet that is traded among users? Thanks again
Hi, thanks for the great question!
Yes, you can compute total profit per user per market using this dataset. Here's how:
Key fields for your use case:
maker / taker: wallet addresses (these are the actual on-chain addresses, consistent with Polymarket's official API)
side: BUY or SELL
price: trade price [0, 1]
token_amount: quantity traded
market_id: to group by market
Approach:
For each wallet address, filter all rows where it appears as either maker or taker
Determine the wallet's direction in each trade: if the wallet is the taker and side=BUY, the wallet is buying; if the wallet is the maker, check maker_direction for their side
Accumulate net token position and cost basis per market
For settled markets (resolved to 0 or 1), the terminal payout = remaining_tokens × outcome_price gives you the final PnL
One caveat: this dataset captures secondary market OrderFilled events only. It does not include token minting/redemption or resolution payouts directly — but since every position must be acquired through trading, you can reconstruct the full traded position from this data. Off-chain transfers between wallets are also not captured, so one user with multiple wallets will appear as separate entities.
Hope this helps! Feel free to ask if you have more questions.
Thanks again for your prompt and detailed response; it's extremely useful. I just want to make sure I understand one point correctly. My concern is that wallet exposure on Polymarket may change not only through OrderFilled trades, but also through token creation and other non-trade events such as split, merge, redeem, and settlement. If that is the case, then a dataset built only from OrderFilled events might not fully capture a user’s true position or final P&L. Could you please confirm whether this is an issue here? Can token creation or other non-trading events generate or close exposure in a way that would not be fully observable from this dataset alone?
Thank you very much
Thanks again for your prompt and detailed response; it's extremely useful. I just want to make sure I understand one point correctly. My concern is that wallet exposure on Polymarket may change not only through OrderFilled trades, but also through token creation and other non-trade events such as split, merge, redeem, and settlement. If that is the case, then a dataset built only from OrderFilled events might not fully capture a user’s true position or final P&L. Could you please confirm whether this is an issue here? Can token creation or other non-trading events generate or close exposure in a way that would not be fully observable from this dataset alone?
Thank you very much
Good question — this is something I was also aware of when putting the dataset together.
You're right that Polymarket's Conditional Token Framework (CTF) on Polygon does have other contract events like PositionSplit, PositionsMerge, and PayoutRedemption that can affect a user's position. So OrderFilled events alone don't tell the complete story.
That said, in practice, I think for most users you can get a fairly good approximation by combining:
OrderFilled events from trades.parquet — for tracking trade-level P&L
Market resolution outcomes from the Polymarket API or market metadata — for handling positions held through settlement (winner pays $1.00, loser pays $0.00)
This should work well for the majority of cases. For full precision — particularly for users who may be doing split/merge operations — you could look into querying the CTF contract events directly on Polygon, where the contract addresses are publicly available.
Hope this is helpful!