Research note 013: memory survives the stress week, modestly stronger
Date: 2026-10-16 · Status: one crisis week, hourly buckets, stated pairs. The stress test notes 011–012 left open: SVB collapse (Mar 10) plus USDC-depeg weekend, ~54M prints at ~7× January's flow. Does imbalance memory strengthen when it matters, and does the session split survive contact with a crisis?
Question
Calm-week memory (r=+0.36) could be a fair-weather artifact — microstructure that evaporates when flow multiplies and regimes break. Same lag-1 method over 2023-03-06..12, overall plus sessions plus magnitude splits.
Dataset and method
- Seven daily spot aggTrades slices, 54,209,013 prints into 168 UTC
hours, overall taker-buy 50.2%. Pins are extracted-CSV SHA-256 in
research/scripts/stress_013.py(stdlib only, one command) — note the discipline explicitly, because this round caught a real slip: zip-level hashes were pasted where content hashes belong, and the script refused them before any math ran. - Identical pairs to note 011 (consecutive hours, same-session only, median magnitude splits) so every number below compares directly.
Results
| Scope | Calm week (011) | Stress week (013) |
|---|---|---|
| All consecutive hours | +0.36 ± 0.08 | +0.49 ± 0.08 |
| Asia 00–08 | +0.34 ± 0.14 | +0.44 ± 0.14 |
| EU 08–16 | +0.34 ± 0.14 | +0.43 ± 0.14 |
| US 16–24 | +0.34 ± 0.14 | +0.60 ± 0.14 |
| High-magnitude first | +0.32 ± 0.11 | +0.60 ± 0.11 |
| Low-magnitude first | +0.03 ± 0.11 | +0.20 ± 0.11 |
Reading
Three facts with descending certainty. First, memory survives 7× flow: +0.49 against 0.08 is unambiguously nonzero — the carryover process does not evaporate under stress. Second, the magnitude gradient replicates and steepens (+0.60/+0.20 vs +0.32/+0.03): conviction hours still carry everything, and in a crisis week even quiet hours lean slightly. Third — stated as unresolved, not as a finding — the session split does not replicate cleanly: the US reads +0.60 against +0.44/+0.43, but the gaps do not clear combined error bars, and neither does the +0.13 overall strengthening (+0.49 vs +0.36 at ~1.1 sigma). "Stronger under stress" is suggestive, not established; what is established is that memory persists at crisis volume with the same shape (magnitude-graded, session-flat within error).
What would strengthen or overturn this: more stress weeks (one crisis is an anecdote with error bars — the honest unit here is still n=167 hours); the perp cut in stress (does balanced flow remember any better when it matters?); maker-side persistence; and the live-captured version, where hours stop being archive buckets.
Reproduce
python3 research/scripts/stress_013.py --data-dir /tmp/stress013
Seven CSV content hashes pinned in the script, refused on any mismatch — verified by a from-scratch rerun with identical figures.