Lucent Composite Fair Value Methodology

Lucent Composite Fair Value Methodology

This page documents how Lucent computes composite fair-value probabilities for prediction-market contracts served under /api/pm/fair-value. The composite is a weighted geometric mean of odds (logit pool) over three provider legs: two external liquid-market anchors (Kalshi, Polymarket) and one internal literature-priors model leg. Published probabilities are decision-support inputs, not investment advice.

Provider set

Three provider legs compose the fair value:

• **Kalshi** — public trade-api v2 price on markets where the registry entry carries a Kalshi anchor. Mid-price from bid/ask when a valid book exists, falling back to last trade. Non-active status maps to freshness 0.
• **Polymarket** — public Gamma API price on markets where the registry entry carries a Polymarket anchor. Preferred price is lastTradePrice; falls back to the YES token in outcomePrices[0]. Closed markets map to freshness 0.
• **Lucent internal model** — literature-derived base rate per (eventType, eventSubtype, therapeuticArea) tuple, plus structured adjustments (CHMP precedent, FAERS safety-signal count). Zero use of biopharma-intel's Hit Rate substrate — that measures alert-capture timeliness, not approval probability.

The composite is a pure logit pool at V1 — EXTREMIZATION_FACTOR = 1.0. Satopää et al. (2014) show extremized logit pool outperforms non-extremized on resolved-outcome corpora with d ∈ [1.16, 3.92], but V1 has zero resolved substrate for these markets; any factor choice would be unjustifiable. V2 revisits once N ≥ 30 resolved markets accumulate.

Weight formula

Each provider leg's raw weight is a product of four factors, all in [0, 1]:

w_i = provider_base_i × freshness_score_i × liquidity_score_i × resolution_match_score_i

Provider base weights: Kalshi 0.4, Polymarket 0.4, internal model 0.2. Raw weights normalize to sum to 1.0 (SPEC D4 Step 1); the model leg is then capped at MODEL_LEG_CAP_WITH_EXTERNAL = 0.4 of the FINAL NORMALIZED weight whenever any external leg is present (SPEC D4 Step 2). Cap redistribution is proportional to external legs' RAW weights:

Given raw per-leg weights w_kalshi, w_poly, w_model:

      // Step 1 — provisional normalize
      total_raw = w_kalshi + w_poly + w_model
      n_i = w_i / total_raw   for each i

      // Step 2 — model-leg cap if any external leg is present
      if (n_kalshi + n_poly > 0) and n_model > MODEL_LEG_CAP_WITH_EXTERNAL:
        excess = n_model - MODEL_LEG_CAP_WITH_EXTERNAL
        n_model = MODEL_LEG_CAP_WITH_EXTERNAL
        external_raw_sum = w_kalshi + w_poly
        n_kalshi += excess × (w_kalshi / external_raw_sum)
        n_poly   += excess × (w_poly   / external_raw_sum)

Worked example 1 (both externals full-quality): raw {kalshi: 0.4, poly: 0.4, model: 0.2}. Provisional n = {0.4, 0.4, 0.2}. Model at 0.2 < 0.4 cap. Final {0.4, 0.4, 0.2}.

Worked example 2 (kalshi unreachable, poly stale + thin): raw {kalshi: 0, poly: 0.16, model: 0.2}. Provisional n = {0, 0.444, 0.556}. Cap fires; excess = 0.156 goes fully to poly (kalshi raw = 0). Final {0, 0.6, 0.4}.

Worked example 3 (both externals absent): raw {kalshi: 0, poly: 0, model: 0.2}. Cap does NOT apply (no external leg present). Final {0, 0, 1.0}. Model leg is *the* answer, not a shrinkage.

Model leg composition

The internal-model leg is a literature-derived base rate plus two structured adjustments. The base rate lives in data/fair-value-model-priors/base-rates.json — one row per unique (eventType, eventSubtype, therapeuticArea) tuple, each carrying {rate, citation, capturedAt, notes}. Sources: FDA approval statistics, BIO/MIT/Tufts CSDD industry meta-analyses, PDUFA priority-review base rates, FDA AdCom calendar reviews, FDA PRV disclosure register, FDA REMS document register. Every prior is version-pinned per SPEC D18: the priors JSON is canonicalized into the methodology-hash sha input, so any edit → sha changes → forces FAIR_VALUE_METHODOLOGY_VERSION bump.

Two adjustments apply on the log-odds axis:

1. **CHMP precedent** — applies when regulator ∈ {fda, both} AND a CHMP opinion exists at event_date ≤ asOf. Positive → +CHMP_POSITIVE_LOGODDS_DELTA (+0.05). Negative → CHMP_NEGATIVE_LOGODDS_DELTA (−0.05). Withdrawn / none → 0. EMA-only markets excluded (would double-count). 2. **FAERS safety-signal count** — applies ONLY when resolutionAxis === "safety" AND count > 0 in the 12 months preceding asOf. Delta = FAERS_PER_SIGNAL_LOGODDS_DELTA × min(count, FAERS_COUNT_CAP) = −0.03 × min(count, 3). Ceiling: 3+ signals saturate at −0.09.

All four coefficient constants (CHMP_POSITIVE_LOGODDS_DELTA, CHMP_NEGATIVE_LOGODDS_DELTA, FAERS_PER_SIGNAL_LOGODDS_DELTA, FAERS_COUNT_CAP) plus FAIR_VALUE_METHODOLOGY_VERSION live in src/fair-value/methodology.ts alongside the weight formula. Editing any of them requires version bump per SPEC D12/D18.

Freshness gate

Per-provider freshness maps quote age to a multiplicative factor (SPEC D6):

• Age < 10 minutes → 1.0
• Age < 1 hour → 0.5
• Age < 24 hours → 0.1
• Older than 24 hours OR unreachable → 0 (leg dropped)

Internal-model leg is always freshness 1.0 (computed at request time). Kalshi + Polymarket endpoints don't expose per-quote timestamps cleanly at the V1 wire shape; a live 200 with a valid active/open market maps to 1.0 and any failure maps to 0.

Availability discipline

The response envelope is a discriminated union on available. Silent-omission over caveated inclusion is the load-bearing prior (Phase 26c.3 canonical).

- available: true — composite computed. Carries fairValue, breakdown, computedAt, methodologyHash, methodologyVersion. - available: false — one of two named reasons: - no_provider_legs_fresh — every provider leg dropped at the freshness gate. - internal_model_missing_prior — the registry entry references a tuple absent from the priors JSON, and no external leg is present to carry the composite.

Unknown marketKey values return 404 at the route layer — NOT available: false. That distinction is load-bearing: available: false is for known markets with insufficient substrate.

Cap on internal model

When any external leg contributes, the internal-model leg caps at 40% of the FINAL NORMALIZED weight (MODEL_LEG_CAP_WITH_EXTERNAL = 0.4). Defers to external liquidity by design; on markets where Kalshi/Polymarket carry real anchors, the model leg is a shrinkage prior. When BOTH externals are absent, no cap fires and the model leg becomes the whole composite. V1 accepts the risk of drowning out genuine model-leg edge; V2 revisits if evidence accumulates.

Snapshot cache

Composite results are persisted append-only to pm_fair_value_snapshots keyed on (market_key, computed_at) with methodology-version + methodology-hash pinning. Requests within 10 minutes of the latest snapshot serve the cached envelope; requests past 10 minutes trigger a fresh compute under a per-market advisory lock. The 10-minute TTL is uniform across statuses (D19) — available: true, available: false, and partial all cache for the same window. Snapshot rows are never updated in place.

Version + drift

The methodology hash is sha256(METHODOLOGY_BODY + "::" + FAIR_VALUE_METHODOLOGY_VERSION + "::" + canonicalize(base-rates.json)). Three inputs, one sha:

• METHODOLOGY_BODY — this prose, byte-exact.
• FAIR_VALUE_METHODOLOGY_VERSION — the version string.
• data/fair-value-model-priors/base-rates.json — read at hash-computation time, canonicalized via the shared sorted-key JSON serializer.

Editing any of the three changes the sha. The methodology-hash regression test pins the current sha. A body edit without a version bump leaves the sha stable — silent-drift risk mitigated only by PR discipline. A priors JSON edit forces the sha to change (D18 closes the priors silent-drift arm). Version-bump discipline mirrors HIT_RATE_COMPUTE_VERSION + BACKTEST_METHODOLOGY_VERSION.

Disclaimer

These composite fair-value probabilities are decision-support inputs, not investment advice. External-provider legs (Kalshi, Polymarket) depend on third-party public APIs whose availability is not part of Lucent's uptime SLA. The internal-model leg is anchored on literature-derived base rates with named citations and structured adjustments — it carries the same known-unknown as any prior-driven estimate on a market with limited resolved-outcome history. Nothing on this page or in the associated API responses constitutes a solicitation, offer, or recommendation to buy or sell any security or take any trading position.