File size: 3,797 Bytes
ebcde1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
# Technical report

> **STATUS: SOURCE-CONTROLLED TEMPLATE — RUN-SPECIFIC REPORTS ARE GENERATED BY CODE**

This is the canonical report location for *Order Flow to Price Impact*. It does
not contain manually copied empirical or synthetic performance results.
`make reproduce-sample` produces a checksum-verified frozen bundle containing its
evidence tier, configuration/input hashes, observed UTC interval and Git state;
`make report` then renders a run-specific document without retraining or
recomputing statistics.

## Research question

When do order-flow imbalance, liquidity, and observable limit-order-book state
predict short-horizon price changes, and how much apparent predictive value
survives fees, latency, uncertain fills, adverse selection, and inventory risk?

## Planned evidence structure

The study separates three questions that are often conflated:

1. Does an observable feature predict a strictly future outcome on an untouched,
   time-ordered test period?
2. Is the effect stable across instruments, horizons, and liquidity or volatility
   regimes, with uncertainty and multiple testing acknowledged?
3. Does any apparent effect survive a separately specified execution model?

The model ladder begins with historical-mean or majority baselines, followed by
transparent linear and regularized models and a bounded tree model. Model choice
uses training and validation data only. Overlapping label intervals require
purging and embargo before final held-out comparison.

## Data and temporal integrity

No data period is hard-coded in this source template. A generated report takes
its actual minimum and maximum event timestamps from the frozen bundle—not merely
from requested configuration dates. Raw observations remain
unchanged; validation findings and any downstream exclusions are recorded
separately.

At decision event `t`, features may use only information proven observable at or
before `t`. Labels begin strictly after `t`. Exchange event time and local receipt
time are distinct where both exist.

## Predictive model quality

No model metrics are embedded in this template. The generated comparison reports
held-out sample size and period, ROC-AUC or regression diagnostics as applicable,
log loss, Brier score, expected calibration error, and serialized fixed-block
bootstrap intervals. Regime-level outcomes and model diagnostics are persisted as
separate descriptive analysis artifacts rather than folded into the comparison
table. Missing metrics will display as `N/A`, never as zero.

## Execution and simulated performance

No execution result is embedded here. Generated simulation results state maker
and taker fees, decision and order latency, queue/fill proxy, partial fills,
adverse selection, inventory limits, liquidation, turnover, and trade-size or
capacity sensitivity. Predictive metrics and execution metrics remain separate.

## Economic interpretation

There is currently no supported claim about predictability, profitability,
statistical significance, or deployability. The synthetic smoke run validates
software plumbing only and carries the mandatory
`SYNTHETIC_SMOKE` warning on every output surface.

## Limitations

See [methodology_limitations.md](methodology_limitations.md) for the maintained
methodology and limitations register. Even a completed public-data analysis will
remain venue-specific, sensitive to timestamp and fill assumptions, and distinct
from realized live execution.

## Reproduction record

Run `make reproduce-sample && make verify-run && make report`. The generated
report records run ID, evidence tier, observed UTC period, configuration SHA-256,
input-manifest SHA-256 values, Git commit, dirty state and runtime metadata. The
source-controlled template intentionally remains free of copied metrics.