live·OBSERVINGautonomyL1 RESEARCHv0.1.0

all phases built — researching real tokens on 22 days of history

experiment #000004

Does a burst of volume mean the pool is still there two hours later?

COMPLETEDINCONCLUSIVENEEDS_MORE_DATA
question

Does a burst of volume mean the pool is still there two hours later?

A measurement where the last hour's volume is at least three times the median of the preceding three hours is followed, two hours later, by liquidity still standing at 80% of its level — more often than a measurement in the same liquidity band with no such burst.

population
Solana tokens under measurement with reported liquidity and hourly volume — the two fields the market source supplies for every pair
sample
Measurements holding at least four readings in the preceding three hours and having a reading two hours later
timeframe
3h of history behind each measurement; the outcome read 2h after it
baseline
Measurements in the same liquidity band where volume did not spike
falsified if
The gap is under 8 percentage points, points the other way, or does not survive being split by liquidity band
dataset
reproducibility
version
token-measurements-v2
hash
b39e6a7452bdd113f02b62cdd2d13ce63c95cc0257f2d8a80cc12878ee9f5fde
sample size
185
train
not applicable (no model fitted)
validation
2026-08-27T04:00:00+00:00 to 2026-08-28T12:45:00+00:00
out of sample
features
volume_usd, liquidity_usd, age_seconds
method

token-measurement cohort comparison. Exposure is the trigger condition evaluated on a 3h trailing window; the outcome is 'liquidity still at 80% of its level' read strictly 2h later — hours of clock time, resolved against the measurement timestamps. Rates are compared pooled and within each liquidity band with a two-proportion z-test, then re-checked on a chronological split.

strata
"liquidity"
window_hours
3
horizon_hours
2
min_effect_pp
8
min_window_points
4
result

Difference of -9.4 points against the predicted direction (exposed 80% vs control 89%, 5 vs 180 measurements). The smallest group holds 5 rows, below the 30 needed to judge a difference of this size, so the falsification rule is not applied to a sample that cannot support it.

effect size
-0.27
p value
0.50
confidence
0.15
z
-0.6709
strata
[object Object]
n_control
180
n_exposed
5
rate_control
0.8944
rate_exposed
0.8
difference_pp
-9.44
excluded_rows
[object Object]
distinct_tokens
28
expected_direction
1
signed_difference_pp
-9.44
first_half_difference
null
second_half_difference
0.0045
sample_too_small_to_judge
true
falsification_condition_met
true
critic
a hypothesis cannot become supported without this gate
NEEDS_MORE_DATA

Smallest group is 5 measurements against a 30 minimum. 999 tokens were excluded for short series and 265 readings had nothing 2h later. Tokens that died early are exactly the ones that run out of series, so the sample leans toward survivors. Exposure is defined by the same detector that surfaced the anomaly, so the hypothesis and its test share a definition. An independent trigger would be stronger. One half of the period had no comparable rows, so stability is untested. five exposed rows spread across at most a handful of the 28 tokens means a single token's fate drives the whole gap, and the control's 89% survival makes the outcome near-universal so it can barely discriminate no model read this result (critic-agent-v1); the deterministic checks stand alone. no model read this result (critic-agent-v1); the deterministic checks stand alone.

version
critic-checks-v1
stability
NEEDS_MORE_DATA
confounding
PASS
sample_size
NEEDS_MORE_DATA
data_leakage
PASS
data_quality
PASS
independence
PASS
agent_version
critic-agent-v1
model_verdict
NOT_REVIEWED
selection_bias
NEEDS_MORE_DATA
look_ahead_bias
PASS
multiple_testing
PASS
survivorship_bias
NEEDS_MORE_DATA
limitations

smallest group has 5 measurements, below the 30 needed to believe a difference of this size

trace #000008
immutable
  1. 08:45:00OBSERVATIONKORI: volume is 3.2x its window median; buy share 25% against a 48% baseline
  2. 08:45:00ANOMALYVOLUME_ACCELERATION score=0.12 [volume-acceleration-v1]
  3. 08:45:00MEMORY_SEARCH20 related memories retrieved before writing the hypothesis (vector-cosine/pgvector)
  4. 08:45:00HYPOTHESIS#4 Does a burst of volume mean the pool is still there two hours later?
  5. 12:50:37DATASET185 measurements over 28 tokens; hash b39e6a7452bd…
  6. 12:50:37EXPERIMENT#000004 volume-burst-pool-2h
  7. 12:50:38CRITICNEEDS_MORE_DATA: Smallest group is 5 measurements against a 30 minimum. 999 tokens were excluded for short series and 265 readings had nothing 2h later. Tokens that died early are exactly the ones that run out of series, so the sample leans toward survivors. Exposure is defined by the same detector that surfaced the anomaly, so the hypothesis and its test share a definition. An independent trigger would be stronger. One half of the period had no comparable rows, so stability is untested.
  8. 12:50:38RESULTDifference of -9.4 points against the predicted direction (exposed 80% vs control 89%, 5 vs 180 measurements). The smallest group holds 5 rows, below the 30 needed to judge a difference of this size, so the falsification rule is not applied to a sample that cannot support it.
  9. 12:50:38MEMORY_UPDATEresult stored: volume-burst-pool-2h: inconclusive