Imported from g8d3/p4 (e025-hyperliquid-candle-tails/ag-03-cond/AGENTS.md). Install upstream with npx skills add g8d3/p4 --skill ag-03-cond. Copyright stays with the author.

ag-03 — Conditional tails: is the extreme predictable?

The core research question: does the next candle's return distribution — especially its extreme tail — change given what just happened? If yes, that is a statistical edge worth building on. If no, that is an honest null result.

Inherits

• ../../e000-fundamentals/AGENTS.md — principles, command rules, data formats
• ../AGENTS.md — experiment scope
• ../ag-02-dist/AGENTS.md — derived columns and their exact definitions

Inputs

• ../ag-01-data/output/candles_raw.csv — raw candles (compute ret and range yourself, same definitions as ag-02, per (coin, tf) ordered by time).

Method

For each signal, split the data into signal-present vs signal-absent, and compare the next-candle return tail (ret[t+1]) of the two groups against the unconditional distribution:

1. Extreme prev move: ret[t] > +2σ, < −2σ, > +3σ, < −3σ (σ = per (coin,tf) stdev)
2. Volatility state: range[t] in top decile, top percentile (volatility clustering test)
3. Volume spike: v[t] in top percentile per (coin,tf)
4. Direction + size: e.g. 5 consecutive up-candles → next distribution

Deliverable table:

output/cond_next.csv
coin | tf | signal | group | n | mean_next | stdev_next | p50_next | p90_next | p99_next | p99.9_next

group = yes (signal present) or no / base (unconditional).

Decision rule

A signal is "interesting" when the conditional tail differs from unconditional in a direction with enough samples to trust it:

• Sample size: n ≥ 300 per group minimum to say anything about p99. Report n; small n = caveat, not a result.
• Shift measure: compare p99_next / p99.9_next of the group vs base, and report the sign (are extremes followed by more extremes — momentum — or smaller moves — reversion?).
• Apply a basic significance check (bootstrap CI on the group's p99, or a Mann-Whitney U on the two groups) — a difference of 0.001% on 50 samples is noise, not an edge.

Deliverables

Honest-results rule

The likely outcome on liquid perps is that close-to-close returns are near random (no edge), and that volatility clustering is the real, measurable effect (extremes cluster in time). Report what the data says. A clean "no predictable edge, but vol clusters" is a successful experiment — it tells us where the strategy should NOT waste capital, and points at range/vol-based work (like e022) as the realistic direction.

Command execution

• Python + pandas (see ag-02 for the venv pattern if pandas is missing).
• Run with timeout + background + self-wake per fundamentals. Verify CSV and report exist before writing done.txt.

Self-command

( sleep 60; tmux send-keys -t 25-3 "Self-wake: check analysis progress. Files produced? errors? done?" Enter ) &

Window: 25-3. On wake, check outputs, iterate, then write done.txt with the headline verdict per signal.

Notify (mandatory)

In addition to writing done.txt, agents MUST notify on completion: notify.sh done "<agent> finished: <headline>" (from ../../e000-fundamentals/bin/notify.sh) On an unrecoverable failure, before giving up: notify.sh error "<agent> failed: <cause>"