Scoring

Overview#

AGDEL uses an R-multiple scoring model that measures reward relative to risk. Each signal specifies four values at creation time:

• entry_price — current market price (set server-side from Hyperliquid mid)
• target_price — where the maker expects price to be at expiry
• stop_price — where the maker's thesis is invalidated (the risk boundary)
• horizon — time until expiry (1m through 24h)

After expiry, the keeper resolves the signal and determines hit/miss. Makers are evaluated by Profit Factor, not hit rate.

Hit / Miss#

A signal is a hit when two conditions are met:

1. Direction correct beyond noise floor — the price moved meaningfully in the predicted direction
2. Non-trivial target spread — the target was far enough from entry to represent a real prediction

signedMove = (resolution_price - entry_price) / entry_price

LONG  (target > entry): hit requires signedMove > noiseFloor
SHORT (target < entry): hit requires signedMove < -noiseFloor

Target spread (|target - entry| / entry) must also exceed the noise floor.

Noise Floor by Horizon#

Moves smaller than the noise floor are indistinguishable from random fluctuation. Calibrated from production data (~p10 of realized absolute moves):

R-Multiple#

R = |target_price - entry_price| / |stop_price - entry_price|

R is capped at 20.

R measures reward relative to risk using the maker's own target and stop prices. A tight stop with a distant target = high R. A wide stop with a close target = low R.

Quality Score#

quality_score = R    (on hit)
quality_score = 0    (on miss)

The quality score is the R-multiple earned. A maker who wins less often but with high R-multiples can be more profitable than a maker who wins frequently with low R.

Profit Factor#

Profit Factor = sum(R on wins) / count(losses)

> 1.0 = profitable maker
= 1.0 = breakeven
< 1.0 = unprofitable

Profit Factor is the primary maker metric on the dashboard. It replaces hit rate because it accounts for the magnitude of wins, not just their frequency. A maker with 30% hit rate but average R of 5.0 has a Profit Factor of 1.5 — profitable despite losing most signals.

Worked Examples#

Example 1 — Hit with R = 3.0:
Entry: $2,000 | Target: $2,060 (LONG) | Stop: $1,980 | Horizon: 1h
Resolution: $2,055

spread = |2060 - 2000| / 2000 = 3.0% > noise floor (0.0244%) ✓
signedMove = +2.75% > noise floor ✓ → direction correct

R = |2060 - 2000| / |1980 - 2000| = 60 / 20 = 3.0
Result: HIT → quality_score = 3.0

Example 2 — Miss (wrong direction):
Entry: $2,000 | Target: $2,060 (LONG) | Stop: $1,980 | Horizon: 1h
Resolution: $1,970

signedMove = -1.5% → direction wrong
Result: MISS → quality_score = 0

Example 3 — Miss (trivial target):
Entry: $2,000 | Target: $2,000.08 (LONG) | Stop: $1,999.95 | Horizon: 1m
Resolution: $2,000.05

spread = 0.004% < noise floor (0.0049%) → trivial target
Result: MISS → quality_score = 0

Example 4 — Profit Factor over 10 signals:
3 hits (R = 3.0, 2.5, 4.0), 7 misses
sum(R on wins) = 9.5, losses = 7
Profit Factor = 9.5 / 7 = 1.36 → profitable (despite 30% hit rate)

Default Penalties#

• No delivery by leg deadline -> purchase-leg default, refund outcome
• No reveal by expiry + 30m -> signal default, refund outcomes

Both defaults result in quality_score = 0 and impact Profit Factor.

Legacy Scoring#

Signals without a stop_price use the legacy points-based model for backward compatibility:

score = ambition × (direction + precision + breakout)
Range: 0–5, hit = score >= 1.0

Legacy signals do not participate in Profit Factor calculations.

Optimization Guidance#

• Set a meaningful stop price at a genuine invalidation level — your stop defines the risk side of R.
• Aim for R >= 2.0 — you can be wrong more than half the time and still be profitable.
• Get direction right beyond the noise floor — direction is binary (hit or miss).
• Don't set trivial targets — if the target spread is below the noise floor, the signal always misses.
• Choose horizons that match your edge — shorter horizons have tighter noise floors but less time for the move to develop.
• Always reveal on time — defaults earn quality_score = 0 and damage your Profit Factor.
• Deliver encrypted payloads promptly — missed deliveries result in refunds.

Aggregation Hierarchy#

1. maker x signal_type x horizon_bucket
2. maker x signal_type
3. maker aggregate (context)

Reliability Adjustment#

Sample size is first-class. A slice with 10 trades should not outrank a stable slice with 1000 trades on raw Profit Factor alone.

reliability_weight = sample_count / (sample_count + k)

adjusted_quality = prior_quality + reliability_weight * (raw_quality - prior_quality)