10 Known-Event Extremes as an Attribution Problem
What Was Known, What Caused the Move, and Which Tail Is Being Estimated
SSRN: abstract_id=7025118 Status: Corrected conceptual working-paper chapter
The earlier hard FAT_BINARY ontology, its 70/8 counts, and the sizing prescriptions said to follow from it are withdrawn. Event knowledge does not imply known direction, magnitude, or probability. A fixed bounded event jump also does not invalidate intermediate-sequence Hill consistency for an unbounded regularly varying structural tail.
10.1 Abstract
Some large equity moves occur near scheduled regulatory, clinical, legal, or corporate decisions. Calling them Black Swans, structural power-law draws, or binary-event draws without documented attribution confuses different questions. This chapter separates what was publicly knowable before the move, what primary evidence attributes the move to, and what statistical model describes the remaining return distribution. At a finite operational threshold, attributed events can affect a tail estimate. That fact does not create a distinct ontological class or imply a universal position-sizing rule. We propose an evidence-preserving attribution layer with timestamped primary sources, competing-event controls, independent adjudication, confidence, and an unresolved state.
10.2 1. Three questions, kept separate
- Event knowledge. Was an event category or decision calendar publicly known before the return occurred?
- Return attribution. Does timestamped primary evidence support attributing the extreme move to that event rather than to another simultaneous cause?
- Tail model. After representing attribution uncertainty, what model adequately describes the return distribution at the threshold of interest?
A yes to the first question does not imply that direction, magnitude, or probabilities were known. A yes to the second does not prove that the non-event process is power-law. A Hill estimate does not answer either attribution question.
This separation is consistent with the distinction between measurable risk and uncertainty (Knight 1921), later discussions of epistemic versus ontological uncertainty (Colander 2010), and robust decision-making under model ambiguity (Hansen and Sargent 2008). It does not claim that those literatures already supply an empirical equity-event classifier.
10.3 2. Correct statistical position
With an unbounded regularly varying structural component and fixed bounded jumps, an intermediate Hill threshold eventually rises above those jumps. Intermediate-sequence estimation can therefore remain consistent. A fixed top fraction can continue to include event observations at operational sample sizes. Unbounded event-size distributions, growing jumps, bounded structural tails, and dependence are different models and require separate analysis.
“Known-event extreme” is consequently metadata about generating context, not an ontological tail class inferred from order statistics. Different event sources may warrant different scenarios or hedges, but no universal capital rule follows from the label alone.
10.4 3. Evidence-preserving attribution protocol
For each candidate extreme, retain:
- event family and pre-event calendar status;
- primary-source URL or document hash and publication timestamp;
- market closes immediately before and after the candidate event;
- competing company, sector, and market events;
- attribution rule and version;
- two independent blinded classifications with confidence;
- unresolved and disputed states; and
- append-only corrections rather than overwritten labels.
The analysis should report inter-rater agreement, class uncertainty, and tail estimates under include, exclude, and probability-weighted attribution. The former 70 structural and eight binary counts remain withdrawn until a protocol of this kind can reproduce them.
10.5 4. Falsifiable research questions
- Does removing high-confidence known-event observations materially change Hill curves at fixed operational bandwidths?
- Does the change vanish under proper intermediate bandwidths as the fixed-jump result predicts?
- Do attributed and unattributed extremes differ in post-event drift, drawdown, recovery, option pricing, or missingness after point-in-time sector and liquidity controls?
- Does attribution add prospective information beyond the full empirical return history?
Null results are informative. If adjudication is unstable or attribution adds no information, TrailMap should preserve the metadata without creating a risk class.
10.6 5. Conclusion
The useful contribution is procedural: document what could be known, what evidence links an event to a move, and which tail estimand is being studied. The former categorical taxonomy, approximately known magnitude/probability language, unaudited counts, and universal prescriptions are not supported.