What my Polymarket trading records showed about profit and risk
The research includes recorded trading, simulated execution, quote-path studies and model tests. Calling all of it a paper test would miss a substantial part of the work. Combining it into one profit number would be just as misleading.
I went back through the written study reports to separate those records. This page concentrates on the documented historical strategy comparison and the accounting questions raised by the later inventory work. The figures below are the reports' documented findings. This review did not rerun the old code, rebuild the ledger or independently verify transaction receipts.
The recorded strategies had different outcomes
The historical strategy audit used a defined decision cohort, excluded recovered inventory and applied its duplicate and write-off rules. A cohort is the group of entries selected for a comparison. These are closed entry rows, not necessarily independent markets. Repeated entries, correlated contracts and changing model revisions remain in the historical groups.
| Strategy | Closed rows | Journal positive and negative rows | Recorded fee-net result |
|---|---|---|---|
| BTC five-minute late favorite | 291 | 279 positive, 12 negative | −$21.03 |
| ETH five-minute late favorite | 166 | 165 positive, 1 negative | +$31.83 |
| BTC fifteen-minute late favorite | 53 | 50 positive, 3 negative | −$1.52 |
| Legacy lower-price fifteen-minute entries | 221 | 96 positive, 125 negative | −$79.45 |
| Static terminal settlement | 34 | 29 positive, 5 negative | −$49.01 |
| Range including two winner-take entries | 75 | 54 positive, 17 negative | −$125.33 |
| Static barrier | 58 | 41 positive, 5 negative | +$81.11 |
Journal signs and fee-net signs are different measures. The positive and negative counts come from journal results; the final column applies the audit's recorded-fee treatment. Zero, missing-result and legacy write-off rows explain why some counts do not add to the closed total. I have not added these groups into a portfolio return.
The ETH result matters. Its documented +$31.83 is evidence against the blanket claim that every variant failed. The static barrier group was also positive in recorded accounting. Neither result should disappear because a separate four-asset paper cohort lost money.
Positive records need their own evidence labels
The late-favorite cohorts had recorded entry fees, and almost all rows linked to confirmed BUY fills. That is materially stronger execution evidence than a chart touch or a displayed offer. It still does not make every recorded fee estimate independently receipt-verified.
The barrier total has a different footing. The report says 53 of its 58 closed rows lacked recorded entry fees and only five linked to confirmed BUY fills. Legacy zero write-offs also affected settlement coverage. Its +$81.11 belongs in the comparison, labelled as recorded accounting rather than certified all-cost profit.
The same distinction applies to negative totals. In the older ordinary range group, 55 of 73 rows lacked entry-fee measurements; 22 of the 34 static terminal rows did too. The range total in the table also includes two winner-take entries. An unknown fee is not a zero fee. Applying today's schedule to an old trade would be an estimate, not a recovered receipt.
Many wins can still leave a negative total
BTC five-minute late-favorite trading recorded $100.12 of fee-net gains against $121.15 of losses. The report described roughly $0.36 per positive result and $10.10 per negative one. A long run of small gains did not pay for the larger losses in that cohort.
ETH recorded $41.87 of gains against $10.04 of losses. That difference is the useful comparison, not a claim that ETH is permanently safer. Asset, period, entry selection, sizing and model version all matter.
The report also showed different results across its historical partitions. Several policy revisions occurred, so a change in the aggregate is not a controlled experiment proving that one release caused the change.
Fixed sizing changed the settlement comparison
The 34 static settlement rows recorded $127.57 of gains and $176.58 of losses. Their three largest losses accounted for about 75% of all recorded losses. That made concentration a concrete research question.
The report rescaled the same historical unit returns to equal $10 all-in budgets. The recorded result changed from −$49.01 to a +$22.21 counterfactual. Range stayed negative under its equal-budget comparison, and the legacy lower-price fifteen-minute group became more negative. Sizing was a useful lead for settlement, not a universal rescue.
The +$22.21 is not an executed fixed-budget result. Linear rescaling does not recreate order quantities, available depth, different fees, concurrent capital commitments or the effect of one position on the next. It isolates the sizing question while holding the historical unit returns fixed.
An order is not a fill and a fill is not settled profit
- Order
An intent or accepted order establishes a request. Unfilled quantity is not inventory.
- Confirmed fill
Matched quantity establishes acquired or sold exposure. Reconcile partial fills and uncertain responses.
- Position and exit
Track remaining shares, entry basis, sales and actual merge or redemption events.
- Economic result
Count each cash event and cost once. Keep open marks separate from closed results.
The admission audits show why these stages matter. One audit documented 445 durable order records but no recorded fills. A separate BTC version audit separately found 187 final-minute scans across 17 resolved markets without reaching an execution attempt. Those are different cutoffs and policies, not contradictory trade totals.
A separate recorded incident showed the opposite problem. An initial attempt found no match, but a later repriced attempt filled and the holding lost at resolution. Reading only the first response would misclassify a funded loss as a missed opportunity. The written replay traced the captured admission decision rather than assuming that faster submission would have prevented the loss.
Polymarket's current order-management documentation likewise separates order state from account trades. Its position documentation distinguishes merging balanced tokens from redeeming winning tokens after resolution. These references explain the mechanics; they do not verify our historical records.
The redemption discrepancy stays visible
The confirmed-breach study used 25 historical entries and compared quoted loss-control actions on a uniform 13-share basis. Its unchanged historical-exit price-and-fee scenario totalled +$4.78. Directly scaling the recorded journal results to that share count gave −$0.04. The report identified redemption rows with an exit price of $1 but near-zero recorded journal P&L.
Those are different measures, and the report did not establish the cause. I cannot label the difference gas, a fee, a proven software bug or extra earned profit. It means the quote scenario must remain a quote scenario rather than become audited portfolio profitability. The documented discrepancy is preserved here instead of silently choosing the more attractive baseline.
Likewise, a local write-off and a later token burn or redemption must not book the same loss twice. Partial-sale cash already included in a trade must not be added again. The historical strategy audit removed two duplicate/write-off rows and explicitly left unmatched legacy partial attribution separate.
Use a capital denominator that matches the claim
Recorded closed-trade P&L is not marked portfolio equity. Open positions still carry risk; an unrealized price change is not a completed sale. A pending redemption is not the same stage as returned collateral. Expected rewards and rebates are not earned cash until the relevant receipt exists.
A return percentage also needs a defined capital base and period. Purchase turnover, the sum of entry costs, peak funded cash and starting portfolio capital are not interchangeable. Overlapping positions tie up capital together. Holding a barrier contract for weeks is different from recycling the same capital several times in a day.
The report totals above therefore remain dollar cohort diagnostics, not a whole-portfolio return or a daily earning rate. The separate paper account has its own virtual starting capital and its own simulated accounting, explained in the paper study.
Keep forecasts and executable trading separate
The research also progressed beyond the earlier API reconstruction. A later direction study documented 17,241 forecasts across 4,316 resolved contracts. Whole expiry episodes stayed together in the split, with outcome-availability purging and an embargo. BTC and XRP improved slightly against their baseline, SOL was unchanged and ETH worsened. All four reported paired uncertainty intervals included zero improvement.
A different frozen return-distribution experiment used synthetic strikes rather than Polymarket executions. Thirteen of sixteen asset-and-horizon cells lost side accuracy; the three small gains had intervals spanning zero. Its historical test had 8,724 daily asset-and-horizon forecasts and 43,620 correlated synthetic labels, not that many trades. This is model research with useful rejection results, not another cash-profit table.
Retrospective threshold work, later chronological checks and future tests have different roles. A candidate that rejects an old losing entry can qualify again later in the same market. A maker quote can be touched without filling. A wider displayed spread can disappear in the cost of completing the unmatched leg. The written studies kept those questions separate; so should the public account.
What this research changed
The useful result is not a single win-rate screenshot. Different strategies had different economics. Sizing materially affected one historical settlement comparison. Execution traces distinguished unavailable offers from orders that really filled. Model studies tested, and sometimes rejected, plausible improvements. Inventory studies exposed the cost of cutting recoveries as well as the cost of holding losses.
These pages explain what we tested, what the evidence showed, and what it means for trading costs, execution and risk.
Study notes and source boundaries
The figures are documented in my historical strategy audit, settlement model audit, confirmed-breach study, admission audits, direction review and return-distribution review. This written-source review covered all 44 Markdown reports and notes in the research-artifacts collection, the notebook documents and the historical strategy snapshot. Ten evidence copies matched the corresponding original Markdown reports.
This was a review of the written methods and results, not a new ledger or code audit. Historical deployment reports describe the versions they checked, not a new runtime test. Private order identifiers and entry settings are not published.
For current venue mechanics, read the fee documentation, order-management documentation and position operations. Current documentation does not establish which fees were actually paid historically.