A7 · Earnings Dealer-Unwind

A paper-tracking research candidate: a name carrying unusually high dealer net-delta into a confirmed earnings session has historically drifted negative on the print.

The short code A7 is display metadata only. The canonical identifier this archetype is keyed by everywhere — config, the screener_candidates.archetype column, the scorer emit — is the descriptive string earnings_dealer_unwind.

Candidate-GENERATOR, not a trade signal. A7 exists ONLY to feed the screener’s forward paper-tracking loop (screener_predictions). It is never a trade recommendation, never a short signal, and never a macro dashboard element. Its dislocation block carries direction-of-evidence as descriptive framing (“historically associated with negative post-event drift, paper-tracking”), not advice.

Honest ABSTAIN-until-history. A7’s edge is placed against each name’s OWN trailing dealer-net-delta history. That history accrues FORWARD from the archetype’s deploy, and the condition needs ≥40 trailing observations per name before it is trustworthy. Until then A7 silently abstains for that name — for weeks after launch, expect zero A7 candidates. That is the design, not a defect.

Thesis

Going into a scheduled earnings print, a name can carry an unusually large net dealer-delta position — the configuration that resolves through a dealer call-unwind once the event lands. Validated on real, calendar-confirmed earnings labels, names whose per-name net dealer-delta exposure sat in their OWN top quintile at the close before a confirmed earnings session went on to drift NEGATIVE over the earnings session itself. The edge is single-session: it lives in the affected session and does not persist a further day.

Direction

PAPER — deliberately neither long nor short. A7 is a candidate-GENERATOR feeding the forward paper-tracking loop, not a trade signal in either direction: “long” would contradict the negative direction-of-evidence, and “short” would dress it up as a short trade signal. The evidence points to negative post-event drift; A7 records that descriptively for paper-tracking and leaves any decision to the operator. On the board it renders as a neutral ◆ chip.

What it looks for

The evaluator (evaluate_earnings_dealer_unwind); thresholds from screener_config.json § archetypes.a7:

• Confirmed next-session earnings (hard gate) — the next session must be a CONFIRMED earnings session for the name, within pre_earnings_days=1 trading days. No confirmed earnings → no emission, regardless of the dealer-delta read.
• Dealer net-delta in its own top quintile — the name’s latest net dealer-delta exposure (dex_net) placed against its OWN trailing history. At or above the quintile=0.8 percentile over ≥ min_periods=40 observations fires PRIMED; the [partial_quantile=0.6, 0.8) band fires PARTIAL. Below min_periods observations → ABSTAIN (the honesty case above).
• Optional confirmers (boost, never required) — per-name net-gamma rank in its bottom quintile; latest price below the per-name gamma flip. Each present confirmer adds to the score; neither is required for either state. (Research demoted both from co-equal signals to optional confirmers.)

The per-name dealer net-delta is Σ(delta × open interest × 100) over the latest option-chain snapshot — calls carry positive delta, puts negative, so the chain’s signed delta orients it. It is derived from the platform’s own option-chain substrate (the per-strike chain already captured for watchlist names) and persisted daily. Vendor-archive evidence motivated the signal; the runtime input is the platform’s own chain substrate, not any external archive.

How it scores

The base score is the dealer-delta quantile (how extreme the position is in the name’s own distribution), lifted by 1 + confirmer_boost × confirmers_fired. The state machine:

• none — no confirmed next-session earnings, no option-chain data for the name (sparse-by-design ABSTAIN like A3/A4/A6), fewer than min_periods trailing observations (ABSTAIN-until-history), or dealer-delta below the partial band.
• partial — confirmed earnings AND dealer-delta in the [0.6, 0.8) band.
• primed — confirmed earnings AND dealer-delta in its OWN top quintile (≥ 0.8) over ≥ min_periods observations.

Paper-tracking only

Every A7 emission lands in screener_candidates and enters the screener_predictions paper-tracking loop exactly like any other archetype, so maturation and grading run on it. Because A7’s edge is SINGLE-SESSION, it paper-tracks at horizon 1 (a per-archetype override in screener_config.json § calibration.per_archetype_horizons) — measuring the affected-session return, the only window the edge was found in. A7 stays uncalibrated until 30+ matured paper-tracked rows accumulate; while uncalibrated, the screener alert gate never fires for it and its sizing tier reads insufficient_data. This is the platform’s standard calibration honesty, inherited for free.

Pitfalls / when it fails

• Single-session only. The edge does not persist past the earnings session; do not read it as a multi-day drift thesis.
• Sparse by design. The per-name chain substrate is watchlist-only; the vast majority of universe names carry no chain data and ABSTAIN.
• Accrual lag. A name needs ≥40 days of its own dealer-delta history before the quintile placement is trustworthy — expect quiet weeks after launch.
• Modest, calendar-confirmed edge. The validated effect is a single ~−0.27 out-of-sample leg with a ~59% sign hit-rate, not a high-conviction directional call — which is exactly why it ships as a paper-tracking candidate-generator and not a trade signal. Its falsifier: 60+ paper-tracked earnings events with a sign hit-rate ≤ 52% kills the archetype.

Related

• The Screener System (vision)
• Screener Pipeline & Lifecycle (operations)
• Sibling archetypes: A1 · Post-Earnings Overreaction · A4 · Pre-Catalyst Coiling · A6 · Vol Mispricing