# SignalsAgent League — Scoring Methodology

**Version 1.0 — effective from first published week. Weekly rebalance.**

## The problem

A signal that makes 500 points on BTC did not do the same work as a signal that
makes 500 points on ES. BTC trades six figures; one percent is over a thousand
points. ES trades four figures; one percent is about 65 points. Raw points are
not comparable across instruments, so a raw-points league is meaningless.

## The rule

**NQ is the base. 1 normalized NQ point = 1 league point.**

Every other instrument gets a weekly conversion ratio:

```
R_i = P_NQ / P_i
```

where `P_NQ` and `P_i` are the **prior Friday's settlement prices** (crypto:
Friday 22:00 UTC print) of NQ and instrument `i`.

A signal's league points for a closed trade:

```
league_points = direction_correct_points × R_i
```

where `direction_correct_points` is signed: points gained are positive, points
lost are negative. Losses count. This league does not hide drawdowns.

## Why this is fair

`R_i = P_NQ / P_i` makes **equal percentage moves score equally**. A 1% move
on NQ, ES, BTC, or DAX all produce the same league points. Percentage return
per unit of price is the cleanest instrument-agnostic measure of a directional
call.

Margin is implicitly handled: exchange margin requirements scale roughly with
contract notional, and notional scales with price. BTC's higher margin and
higher point count cancel out under price normalization. We considered an
explicit margin-ratio adjustment and rejected it: margins change midweek at
the exchange's discretion, differ across brokers, and would make the league
un-auditable. Price normalization is public, reproducible, and disputable by
anyone with a chart.

## Rebalance cadence

**Weekly.** Ratios are computed from Friday settlement and fixed for
Monday 00:00 UTC through Sunday 23:59:59 UTC. Published every Monday on
`/league` and in the API (`GET /api/league/ratios`) before the session opens.
A trade is scored with the ratio table of the week in which it **closes**.

Why weekly, not daily: daily ratios are noisier, invite disputes, and let a
provider game entry timing around the daily reset. Weekly is stable, and any
drift within a week is symmetric — it helps and hurts everyone equally.

## Example ratio table (seed values — recomputed weekly)

| Instrument | Reference price | Ratio R_i | 100 raw pts → league pts |
|-----------|----------------|-----------|--------------------------|
| NQ        | 23,500         | 1.0000    | 100.0 |
| ES        | 6,450          | 3.6434    | 364.3 |
| DAX       | 25,000         | 0.9400    | 94.0  |
| NKD       | 41,000         | 0.5732    | 57.3  |
| BTC       | 118,000        | 0.1992    | 19.9  |
| GC (Gold) | 3,350          | 7.0149    | 701.5 |

Check: a 1% move = 235 NQ pts × 1.0 = 235; = 64.5 ES pts × 3.6434 ≈ 235;
= 1,180 BTC pts × 0.1992 ≈ 235. Equal work, equal score.

## League standings

- Season = calendar quarter. Weekly table published, quarterly winner crowned.
- Score = sum of league points over all **closed** trades in season.
- Also published per signal: trade count, win rate, average league points per
  trade, max drawdown in league points. One number never tells the story.
- Minimum 10 closed trades per week to appear in the ranked table (below the
  cut you're listed as "unranked — insufficient sample").
- The evaluator is agnostic about who owns the signal. Longfort's own signals
  are scored by the same code path as third-party signals. No exceptions, no
  overrides. Scoring code is published.

## Evaluation tape & the Friday rule

Results are produced by Umpire, the deterministic paper executor
(`docs/PAPER_EXECUTOR_SPEC.md`), judging every signal — house and third-party —
on **official exchange prints** (CME/Eurex via Databento/Rithmic; Kraken WS for
crypto). Broker feeds are not evaluation sources: they differ by broker and
cannot be independently verified. Only whitelisted, independently checkable
symbols are listed.

Open-ended signals (no stop/target) stay open until the generator's own exit
command; absent one, they are **force-closed at the first print at or after
Friday 20:00 UTC** of their opening week (`auto_closed_eow`). Unrealized P&L
never counts. No zombie positions.

## Anti-gaming provisions

- Signals are scored on **published** entries/exits only (timestamped when
  the signal fires through the platform, not backfilled).
- No retroactive edits. A published signal is immutable.
- A provider may retire a signal; its history stays on the board for the season.
- New signals enter mid-season with zero; no imported track records count
  toward the league (you can display them, clearly labeled "unverified,
  provider-supplied").

## League epoch & pre-platform records (policy set 2026-07-19)

The official league starts at the **league epoch: 2026-07-19**. Only trades
closed on or after the epoch AND published through the platform count in
official standings — for everyone, house included.

At launch, the house imported its own pre-platform history (real broker fills,
verified against IG execution records and QuestDB OHLC). Per the rule above,
**none of it is league-scored.** It is displayed in a separate, clearly
labeled "Pre-platform records" section for transparency: you can see how house
signals behaved before the league existed, including the ugly ones, but the
scoreboard starts at zero. Provenance is enforced in code: any result closed
before the epoch is automatically classified `imported` regardless of who
pushes it or what they claim.

## Governance

Ratio table generation is `scripts/update_ratios.js` — deterministic from
Friday settlement inputs, committed weekly. Anyone can recompute it.
Disputes: recompute it. If our number is wrong we fix it and say so.