Strategy Health

Closed-trade P&L sliced by the dimensions most likely to reveal a real edge or a real leak, plus what's being filtered out before it ever becomes a trade. Machine-readable copy of this page: /strategy.json. Threshold retuning proposals: latest backtest report.

By Time to Resolution

How close to market close the trade was opened. Entering near expiry has been the single largest leak in this system — near-close trades leave no room for Kalshi to converge before the market resolves.

By Wallet Score

Classifier score of the wallet whose trade fired the signal. Watch for variance *within* a bucket — check the wallet table below before setting a score floor.

By Market Type

By Side

By Gap Size at Signal

Gap relative to the Kalshi price — which is also the return on capital if the gap fully closes, since cost is entry price × size. Read against the entry-price panel below: the two are strongly related, because a fixed 3¢ gap is a large relative edge on a cheap contract and a small one on an expensive contract.

By Entry Price

The strongest single lever in the book so far, and largely what the gap panel above is proxying. It also sets what the fixed 10¢ stop-loss means in risk terms: 10¢ is ~68% of a 15¢ entry but ~13% of an 80¢ one, so the same stop is near-unreachable at the cheap end and a hair-trigger in the middle.

Worst Tickers

Minimum 10 closed trades, worst P&L first.

Worst Wallets

Minimum 10 closed trades, worst P&L first. A wallet losing consistently here is a blocklist candidate regardless of its classifier score.

Settlement-Mismatch Suspects

Pairs whose Kalshi/Polymarket gap was still wide when the trade expired. A gap that never closes suggests the two markets don't resolve on the same condition — a matching problem, not a slow signal. Populates going forward only (194 expiries recorded so far).

Expired Book

Positions abandoned at the max-age timeout rather than closed, so they realized nothing and pnl stays NULL. Two different questions: mark-to-market is what the position was worth when we walked away — known for every expired trade, since the exit price is always recorded. Settlement is what it would have paid had we held on — knowable only once the market actually resolves, which for markets running into 2027 is a long way off. A large gap between the two columns means the timeout itself is costing (or saving) money.

Shadow Filters

Filters that are armed but not blocking. Each row is what the filter would have stopped, alongside the P&L of the trades those signals actually became — so a negative figure is money the filter would have saved. This measures a proposed rule against live data before it changes any behaviour. Enable one by flipping its constant in detector/propagation.py (e.g. NEAR_CLOSE_FILTER_ENABLED). Filters overlap — one candidate can trip several — so each row answers "what would this filter have done on its own?" and the rows do not sum to a total.

Shadow Stop-Loss (relative)

The live stop is a flat STOP_LOSS_CENTS = 10¢, which is ~68% of a 15¢ entry but ~13% of an 80¢ one — near-unreachable at the cheap end, a hair-trigger in the middle. This records where a SHADOW_STOP_PCT-of-entry stop would have exited instead. Exits can't be backtested (the replayer only re-decides entries against realized exits), so this is measured forward on live trades. Nothing here changes an exit.

A positive delta means the relative stop would have done better than what actually happened. Only finished trades count — an open position has no realized side to compare against.

Threshold Stability Across Windows

A single backtest report can't justify a threshold change: the top candidate by validation P&L is exactly the config most likely to have been flattered by noise in that one window. This counts how often each config beat live on both train and validation while keeping ≥70% of live's trades. Read qualified as a fraction — 1/1 is noise, 4/4 is a finding — and check worst: a config that ever went negative is not a candidate, however good its average. Windows evaluated: 2026-07-12, 2026-07-19, 2026-07-26, 2026-08-06.

Nothing here is applied automatically. Changing a threshold means editing the constants in detector/propagation.py by hand.

Layer-1 Candidate Funnel

Every candidate that cleared the cheap structural checks (tracked pair, min trade price), regardless of whether it became a real trade. 'success' means it opened a paper trade. Not era-filtered — logging began when signal_candidates shipped.