Event Linkage Arbitrage

The Event Object

{
  "event_type": "tariff | sanction | merger | litigation | regulation | policy",
  "surprise": -2 to 2,
  "duration": "one_day | multi_week | structural",
  "primary_assets": ["list"],
  "secondary_assets": ["list"],
  "factors": ["inflation", "growth", "liquidity", "regulatory"],
  "confidence": 0 to 1
}

The market is flooded with text. The value of GPT-style extraction is not that it predicts the terminal price response. The value is that it standardizes heterogeneous documents into a single event object fast enough for downstream ranking, basket construction, and execution logic.

Three Tradable Buckets

• Sympathy continuation: second-order names that have not moved enough yet
• Overreaction fade: low-quality headline moves without durable cash-flow impact
• Regime repricing: persistent policy or regulatory shifts that justify a slower multi-day rebalance

Most retail traders lump all event trades together. That is a mistake. A tariff headline and an FDA approval are both "news," but they produce completely different decay profiles and execution demands. The event object needs a duration field specifically so the portfolio knows whether to fade, press, or scale in slowly.

Pipeline Design

1. Ingest wires, press releases, filings, speeches, and transcripts.
2. Deduplicate documents to avoid counting the same event multiple times.
3. Extract event type, affected entities, expected duration, and likely economic channels.
4. Map entities into sector and factor exposures.
5. Compare actual market move versus model-implied move.
6. Trade the laggard basket or fade the overshoot.

If duration = structural and confidence > 0.75:
    trade slower, smaller, multi-day

If duration = one_day and confidence < 0.55:
    look for fade, not continuation

If secondary basket has not moved but primary basket has:
    trade the lag, hedge market beta

Practical Playbooks

Policy and Regulatory Shocks

These often matter more for sector baskets and factor spreads than for single-name trades. If a regulatory headline alters reimbursement, capital rules, export restrictions, or tax treatment, the first-order impact can be obvious while the cross-sector read-through is slower.

Litigation and Legal Rulings

These are dangerous for naive text systems because legal nuance matters. The model should identify event type, likely appeal path, timing, and peer exposures, but high-consequence legal trades should retain manual review.

Geopolitical and Trade Shocks

These are ideal for event linkage because the primary impacted names often move instantly while suppliers, substitutes, freight providers, regional ETFs, and commodity-linked assets move more slowly and inconsistently.

M&A and Corporate Actions

GPT extraction can help identify breakup risk, regulatory sensitivity, financing path, and sympathy read-through to peers. It should not replace the hard work of merger arb spread analysis, but it can improve screening and peer mapping.

Execution and Timing

Event alpha decays quickly, so playbooks matter.

• Immediate window: classify event and map first- and second-order exposures
• First hour: decide whether the move is underreaction, appropriate reaction, or overshoot
• Day 1 to day 5: harvest lagged repricing in secondary names or sector spreads
• Multi-week: only for structural events with confirmed duration and broad economic impact

The event object should feed a basket engine, not a stream of impulsive single-stock trades.

Big Risks

• News duplication: the same event may hit many wires and inflate confidence falsely
• Legal nuance: litigation and policy text are subtle; keep a manual review layer
• Crowded themes: obvious sympathy trades get front-run instantly in mega-caps
• Execution: event alpha dies fast; use liquid names and predefined playbooks
• Duration mismatch: fading a structural event or pressing a one-day event is a common error