Signal methodology and Signal Transparency Lab metrics

Code: src/lib/signals/engine.ts (rules + lifecycle), src/lib/signals/store.ts (persistence), src/lib/signals/metrics.ts (metrics).

Classification taxonomy

Every signal row carries exactly one classification (enforced by a CHECK constraint):

Metrics can be broken down by classification (byClassification); never mix them without saying so.

Integrity rules

• signals rows are immutable: database triggers reject UPDATE and DELETE. A signal is written once, at generation time, with its entry, stop, targets, entry conditions, closure criteria, data sources, timeframe and market regime.
• Lifecycle is the append-only signal_events log, one row per state per signal (opened, tp1, tp2, stopped, trail_exit, expired, closed, invalidated). The current state is the latest event.
• Admin corrections go to the append-only signal_corrections table (field, old value, new value, reason, actor, time). The original row is never changed; corrections are shown next to it. Metrics are computed from the original record.
• Losing signals are never deleted, hidden or filtered. Imports bring in losses exactly like winners.

Strategy sky-capitulation version 1.0.0 (paper)

Parameters are stored in signal_strategies (engine, version, params JSON). The engine refuses to run if the stored parameters for a version differ from the code: any parameter change requires a new version.

Universe: Binance USD-M perpetuals, the top 120 by 24h quote volume with at least $20M volume, from the latest market snapshot (must be < 15 min old).

Trigger — evaluated once per hour on the most recently closed 1h candle:

1. Close is at least 15 % below the highest high of the prior 24 candles.
2. Candle quote volume is at least 3× the median quote volume of the prior 24 candles.
3. Lower wick (min(open, close) − low) is at least 50 % of the candle range (high − low).
4. No open signal from this strategy on the symbol.
5. No new signal on a symbol within 72 h of a stopped event on that symbol (any stop, including a breakeven one).

The scan runs only within 15 minutes after the candle closes; a missed hour is not backfilled (no retroactive signals). Signal ids are deterministic (engine|version|symbol|candle time), so re-running is idempotent.

Market regime at generation: BTCUSDT 24h change > +3 % → risk_on, < −3 % → risk_off, otherwise neutral (unknown if unavailable). Timeframe is 1h.

Plan (long only):

Note: once TP2 is reached the trail level is at least 1.20 × 0.92 = 1.104 × entry, so the trail always dominates the breakeven stop in this version.

Lifecycle resolution (advanceOpenSignals): 5m Binance candles from the last state change onward are walked in time order. Only closed candles are used. Rules inside one candle:

• The stop is checked first. If a candle touches both the stop and a target, the stop is assumed to have hit first (conservative).
• A candle that opens through the stop fills at its open (gap), not at the stop.
• Targets fill at the target price. TP1 and TP2 may hit in the same candle.
• The trailing stop is enforced from the candle after the TP2 candle; the TP2 candle's high seeds the trail.
• If no market data arrives for 24 h after expiry, the signal is marked invalidated and excluded from metrics with that reason.

Event timestamps of tp1/tp2/exits are the open time of the 5m candle where the level was touched.

Costs: 0.05 % fee + 0.05 % slippage per side on every fill (entry, each partial exit, final exit), as a percentage of that fill's notional relative to the entry notional:

gross%     = Σ fraction_i × (exit_i / entry − 1) × 100
sides      = 1 + Σ fraction_i × exit_i / entry
fees%      = 0.05 × sides
slippage%  = 0.05 × sides
net%       = gross% − fees% − slippage%

The result is stored on the closed event with gross, fees, slippage, fills, exit reason and holding time.

Caveat on entry: the reference entry is the 1h close. The signal becomes visible a few minutes later (at generation). A reader acting on it would usually get a different price.

Opens and closes are published to Sky Feed (category = signal, source = signals, entitlement = premium, dedupe key signal:{id}:{state}). The engine is disabled by the signals.engine flag.

Metric definitions (computeMetrics)

All P&L metrics are equal-weighted (each signal = one unit of notional), unleveraged (percent of entry notional, not of margin) and net of fees and slippage as recorded in each signal's result.

Exclusions are surfaced in excluded: [{ id, reason }], never dropped silently:

• alert-only records (signal_type alert) — no defined entry/exit methodology;
• missing entry price;
• invalidated signals (with their reason);
• closed without a realized result, or without a close time;
• terminal exit state without a closed result.

What is NOT claimed

• No independent verification. Results are computed by Sky Terminal from public exchange candles and its own records. No third party has audited them.
• Not executed trades. paper signals are not orders. Real fills, partial fills, queue position, funding payments, liquidation risk, outages and latency are not modelled beyond the flat fee/slippage assumption.
• Imported results are as reported by the source and may be gross of fees.
• Past results do not predict future results. A small sample can look very good or very bad by chance.
• Nothing here is investment advice or a recommendation to trade.