Stage 5

Fair-value engine

Turn the shared feature vector into a calibrated probability and a sensible uncertainty band.

Challenge we are solving

A point estimate p_hat with no uncertainty is dangerous. The same p_hat=0.67 with a 0.04 band is a different bet than with a 0.20 band.

What this stage does

Reads the feature vector, runs the scoring model, and emits p_hat, lower bound, upper bound, an uncertainty buffer, and a suppression reason if any precondition failed.

Why this stage exists

The next stage (execution) and the gate after it (risk) both need to know the *uncertainty*, not just the central estimate.

Flow

FeatureVectorfrom stage 4

Modelv=0.7.2

Calibrationisotonic

Outputp_hat · LB · UB

Suppression?if features incomplete

What the backend should expose

score_id, p_hat (0..1)
LB, UB (confidence band)
uncertainty_buffer (used by decision stage)
model_version, calibration_version
feature_completeness % (input quality)
suppression_reason (null when scoring proceeded)

Maths we expect here

Every formula below is implemented in packages/polytraders-bots/ or packages/polytraders-runner/. Treat the worked example as the unit-test sanity check you should be able to reproduce locally.

Point estimate p_hat

\[\hat p = \sigma(w^{\top} x + b),\qquad \sigma(z) = \tfrac{1}{1 + e^{-z}}\]

Symbol	Meaning	Units / range
\(x\)	Feature vector from stage 4	R^n
\(w, b\)	Model weights and bias	R^n, R
\(\hat p\)	Predicted probability YES resolves true	0..1

worked example\[w^{\top}x + b = 0.71 \;\Rightarrow\; \hat p = \sigma(0.71) = 0.670\]

We do not trade \hat p alone. The uncertainty band below is just as important.

Confidence band (calibrated quantiles)

\[\text{LB} = Q_{0.10}(\hat p \mid x),\;\;\text{UB} = Q_{0.90}(\hat p \mid x)\]

Symbol	Meaning	Units / range
\(LB\)	10th-percentile of the calibrated posterior	0..1
\(UB\)	90th-percentile of the calibrated posterior	0..1

worked example\[\hat p=0.670,\;\text{LB}=0.64,\;\text{UB}=0.71\]

Calibration is isotonic regression refit weekly on resolved markets. Wider bands → stage 7 cuts size.

Uncertainty buffer

\[buffer = \alpha \cdot (\text{UB} - \text{LB}) + \beta \cdot drift\]

Symbol	Meaning	Units / range
\(\alpha\)	Width-to-buffer multiplier	default 0.5
\(\beta\)	Drift-to-buffer multiplier	default 1.0
\(drift\)	Recent calibration drift vs. baseline	0..1

worked example\[(\text{UB}-\text{LB})=0.07,\; drift=0.01 \;\Rightarrow\; buffer = 0.5{\cdot}0.07 + 1.0{\cdot}0.01 = 0.045\]

Stage 7 subtracts this from net_edge before approving a trade. Bigger band ⇒ bigger buffer ⇒ less likely to trade.

Brier score (model quality, backtest)

\[\mathrm{Brier} = \tfrac{1}{N}\sum_{i=1}^{N}\big(\hat p_i - y_i\big)^2\]

Symbol	Meaning	Units / range
\(\hat p_i\)	Predicted probability for market i	0..1
\(y_i\)	Realised outcome (0 or 1)	{0,1}
\(N\)	Number of resolved markets in sample

worked example\[N=500,\; \mathrm{Brier}=0.181 \;\Rightarrow\; \text{better than uniform 0.25}\]

Stage 10 reports Brier per model_version. Promotion threshold: Brier ≤ 0.20 over the rolling 500-resolution sample.

How a developer codes this stage

Reference TypeScript implementation lives in packages/polytraders-* at the repository root. Stage owners maintain these files — read them before writing new code.

packages/polytraders-bots/src/strategyStrategy bots that produce p_hat. Reference: makertight, crossmarketarb.
packages/polytraders-contracts/src/OrderIntent.tsThe OrderIntent carries p_hat, LB, UB, buffer — it is the artefact this stage hands to execution.
packages/polytraders-backtest/src/engine.jsBacktest harness that scores model_version on historical event log → Brier.

See it in the platform mock

The platform mock is the source of truth for what each stage's UI exposes. Open these alongside the code references.

↗Live fair-value view with p_hat, LB, UB/mock-app/fair-value.html ↗Explainability for a single score/mock-app/decisions-explain.html

Reason codes emitted at this stage

STRAT_*STRAT — strategy, model, fair-value

Hover or tap any reason code on this page (or anywhere on the site) to see its canonical short description. Full registry: /standards/reason-codes.