BL-SD-033: Session Structure in Spread

Status: RELEASED

Classification: PUBLIC

Revision: 1

Date: 2026-09-08

Venue: Hyperliquid perpetuals Programme: I Verdict: FALSIFIED Axis of death: Idea Transferability: UNIVERSAL

1. Identification

This dossier tests a single, narrow claim: whether the cost of trading on Hyperliquid varies by the hour of day, in a way that would give an operator a cheap window to trade in. It is written as an observation, not a strategy: no forward return is touched anywhere in this measurement.

2. Economic thesis

If half-spread or impact-spread is structurally thinner during a particular band of UTC hours, an operator could route execution to that band and extract a real cost saving, which would matter directly for a cost-bound hypothesis like BL-SD-032. There is no counterparty forced to lose in this observation; it is a question about the venue's own mechanics, not a trade. The thesis being tested is a piece of trading folklore repeated widely enough to merit direct measurement: that liquidity is thin and manipulation common between 00:00 and 06:00 UTC, and that a cost-sensitive strategy should avoid or favour that window accordingly.

3. Falsifiable claim

Pre-registered before the instrument was run:

Half-spread and impact-spread differ structurally by UTC hour. Structure is confirmed by a dispersion ratio, the 90th over the 10th percentile of the hourly medians, that is materially above one and consistent across at least two of BTC, ETH, and SOL; a ratio near one falsifies the claim for that metric.

4. Mathematical specification

Notation per BL-D-011.
 
Bucket assignment:
 
  utc_hour(t)         = floor(t_ms / 3,600,000) mod 24
  session_of_hour(h)  = ASIA (0-8 UTC), EU (7-16 UTC), US (13-21 UTC), overlap by design
 
Per-bucket metrics, from book and assetctx state observed at t:
 
  half_spread_bps(t)   = (ask0(t) - bid0(t)) / 2 / mid(t) * 1e4
  impact_spread_bps(t) = (impact_ask(t) - impact_bid(t)) / mid(t) * 1e4
  depth_top5_usd(t)    = sum over levels k=1..5 of ( px_bid_k*sz_bid_k + px_ask_k*sz_ask_k )
 
Dispersion statistic, per metric m and coin c, over hourly medians h = 0..23:
 
  dispersion_ratio(m, c) = quantile_90{ median(m, c, h) } / quantile_10{ median(m, c, h) }

Causality holds by construction: every metric is computed from book and context state observed at the same instant t it is bucketed into, with no forward reference. There is no signal, entry, exit, sizing, or cost rule in this specification beyond the metrics themselves, because this measurement characterises the venue's own cost surface; it never places a trade. The pre-registration explicitly prohibits computing markout, adverse selection, or any forward-return quantity at this stage.

5. Operating envelope

6. Instrument

BL-IS-012, test-driven with 12 of 12 unit tests passing, full suite run.

7. Results

Half-spread held constant across all 24 hours on all three coins tested: approximately 0.079 basis points on BTC, 0.28 on ETH, and 0.064 on SOL, every hour, with a p90-to-p10 ratio of 1.01 on each. The folklore claim under test, that liquidity thins and manipulation concentrates between 00:00 and 06:00 UTC, is directly contradicted: spread does not widen in that window, and realised volatility and trade rate are at their lowest there rather than elevated, which is the opposite of what an elevated-manipulation reading would predict.

8. Verdict, axis of death, and transferability

Verdict: FALSIFIED. Axis: idea. Transferability: UNIVERSAL.

The idea under test, that trading cost varies structurally enough by hour to give an operator a cheap window, fails cleanly and consistently across all three coins tested. Half-spread is mechanically tight around the clock on this venue; there is no cheap hour to route toward. No envelope substitution changes this: it is a fact about how this venue's book behaves, not a parameter of any operator's cost schedule, which places it in the UNIVERSAL class.

The finding closes one specific lever for BL-SD-032 directly: the cost gap that hypothesis needed to close could not be closed by scheduling trades to a cheaper hour, because no hour is cheaper. What does vary by hour is activity, not price: trade rate and realised volatility concentrate in the US session, roughly two to three times the trough seen in the Asian early-morning hours, but that is a statement about how much happens, not about how much it costs to participate.

9. Cost and capacity analysis

Not applicable to a characterisation. The half-spread figures reported in section 7 are the relevant cost facts; there is no separate PnL or capacity computation here.

10. Envelope sensitivity and reopening conditions

No binding parameter exists to name, because the underlying metric shows no structure to substitute a parameter against.

Reversal condition:  none computable.

A dispersion ratio of 1.01 leaves no room for a cost-based reversal argument. There is no hour at which half-spread is materially thinner, so no envelope substitution based on timing rescues a cost-bound hypothesis by this route.

What would reopen it: a different metric or a different venue where spread is not mechanically arbitraged flat around the clock. Not tested here.

What would not reopen it: re-examining this same venue's spread by weekday-versus-weekend split. The result was flat enough on the UTC-hour axis that this split was judged immaterial to the finding and was not run.

11. Reference implementation

BL-IS-012 is the specification, and the instrument is published as 05_INSTRUMENTS/strategies/python/src/session_structure.py with 52 tests, shared with BL-SD-034 and BL-SD-035. The data-lake reading it sits on is not reproduced.

Two design choices in the published module bear on how this row should be read.

The session map is fixed in advance and its bands overlap, because real trading hours do. An hour inside two bands is resolved in a declared order, US then Europe then Asia, and the late hours are labelled OFF rather than folded into the nearest session. The order is not a claim about which session matters more; it is a claim that a label has to be chosen before any measurement, so the partition cannot be adjusted after seeing which cut separates best.

Dispersion is reported three ways, and that is deliberate rather than redundant. A robust percentile ratio, the full range, and a coefficient of variation disagree with each other exactly when a single hour is unusual. This row's question is whether quoted width varies enough across the day to matter, so reporting only the robust measure would hide such an hour and reporting only the range would let one thin hour decide the finding. A test constructs that configuration and pins both readings.

12. Evidence register

This measurement is recorded in a narrative report backed by the instrument's own reproducible output, and is registered as D-026 in BL-A-006.

13. Lessons extracted

This dossier corrects a specific piece of trading folklore with this venue's own data rather than accepting or repeating it uncritically: the claim that low-liquidity overnight hours carry elevated manipulation risk does not hold on Hyperliquid, where spread is mechanically flat around the clock and the quiet hours are quiet in activity, not in cost. Filtering an inherited claim against measured data before it is allowed to shape a design is the same discipline this programme applies to its own hypotheses, applied here to an assumption imported from outside the programme.

14. Cross-references

BL-SD-032 for the cost-bound hypothesis this measurement closes one reopening path for. BL-SD-034 for the companion measurement on event supply by session. BL-SD-035 for the companion measurement on reversion quality by session. Ledger row TOC1.

Copyright (c) 2026 Bijective Labs. All rights reserved.