Ingyu Koh · Applied statistics, microstructure and execution research
Loading public report…
| Policy | Net replay P&L | Filled contracts | Drawdown | Day-block 95% interval |
|---|
Inventory persists across quotes until real contract settlement. Volume caps partial fills. A 20-contract net position limit and inventory price skew apply. Hindsight selects the best fixed action for each market using future outcomes; it is unattainable and not a universal upper bound.
These labels are inferred hypothetical fills from actual trades, not recorded own-order fills. Gradient boosting and logistic share the same inputs and chronological split. Quantity fraction and adverse selection are separately modeled.
Choose an example and calculate.
Historical examples use an opening strike anchor between Binance and BRTI. It is not a measured contemporaneous index correction. Editing spot or strike switches to unadjusted custom inputs. Fair value is a spot-digital approximation; the actual settlement averages the final 60 seconds of BRTI. Predictions are not calibrated confidence.
LinUCB is trained with delayed settlement rewards from replay and frozen before evaluation. It is not evidence of live randomized exploration or counterfactual reward identification.
Download reproducible source · Full report · Source-linked examples · Health · Kalshi data documentation
No credentials, client contracts, or live orders. Request bodies excluded from logs. Public bounded requests: 20/minute per runtime; not a distributed spending cap.