Microstructure/docs/PROJECT_PLAN.md

Project plan

Objective

Answer, with defensible event-time evidence, when order-flow imbalance and liquidity conditions predict short-horizon price movement, and how much apparent value remains after fees, latency, fill uncertainty, adverse selection, and inventory constraints. The system is research-only and has no live-order path.

Delivery strategy

The first deliverable is one narrow, fully reproducible vertical slice. Broader market coverage and more sophisticated models follow only after its contracts, timing, and execution accounting are tested.

Milestone	State	Depends on	Deliverable	Acceptance evidence
M0 Repository contract	Complete	None	Instructions, plan, decisions, status, packaging	Required files exist; Python 3.12 environment installed; config/provenance tests, Ruff, and mypy pass
M1 Event data foundation	Complete	M0	Binance trade adapter, optional L2 collector, schemas, manifest/checksum, partitioned Parquet	Mocked retries/metadata/pagination; UTC/schema and content-addressed storage tests pass
M2 Book and quality controls	Complete	M1	Snapshot/delta replay, sequence validation, quality findings	Gap, overlap, crossed-book, duplicate, ordering, invalid value, spread, silence, and clock tests pass
M3 Leakage-safe dataset	Complete	M1-M2	Trade-flow/book features and future labels	Causal lineage, future-label, censoring, and gap-isolation tests pass
M4 Model evaluation	Complete	M3	Baseline, logistic/regularized linear, tree; purged walk-forward evaluation	Prior/unpenalized/L2/tree ladder, calibration, bootstrap, purge, and OOT tests pass
M5 Execution research	Complete	M3-M4	Market/limit fills, latency, partial fills, costs, inventory, liquidation, sensitivity	Deterministic accounting, OOS, volume-conservation, continuity, partial-fill, and liquidation tests pass
M6 Reproducible vertical slice	Complete	M1-M5	One CLI-driven sample run and machine-readable artifacts	Atomic producer and corruption/idempotence tests pass; canonical smoke requires no network
M7 Research communication	Complete	M6	Generated technical report/table, IC memo, limitations, dashboard, portfolio material	Render/dashboard tests read only verified artifacts and visibly label evidence tier
M8 Broader empirical study	In progress	M6-M7	Closed trade-only study plus complete software for the replacement four-date live-L2 capture/analysis campaign	Canonical trade and superseded L2-v1 authorities remain immutable; capture Aug 10--13 under one new campaign authority, publish the Aug 10/11 LOCKED or NOT_CREATED development authority before Aug 12, and publish a recursively verified result or protocol-declared insufficiency

First vertical slice

small documented event fixture
  -> normalized partitioned Parquet + manifest
  -> explicit data-quality findings
  -> causal trade-flow / liquidity-state features
  -> strictly future return and direction labels
  -> purged walk-forward baseline, logistic, and tree models
  -> latency- and cost-aware simulation
  -> JSON/CSV/Markdown run artifacts and dashboard inputs

The offline fixture is an explicitly synthetic smoke test for software reproducibility, not empirical market evidence. make download-sample adds a small public-data path without making the test suite depend on network access.

Dependencies

• Python 3.12; Polars/PyArrow for bounded-memory transformation and Parquet.
• DuckDB for partition inspection and aggregation.
• scikit-learn for transparent, CPU-friendly models and calibration.
• pytest, Ruff, and mypy for verification.
• Streamlit for a local read-only research dashboard.
• Public Binance market-data endpoints only; no API key or account connection.

Risks and mitigations

Risk	Consequence	Mitigation / acceptance test
Snapshot/delta mismatch or sequence gap	Corrupt book state and false imbalance	Halt the affected segment, emit a finding, require a new snapshot; replay gap tests
Exchange timestamp is not local observability	Optimistic latency and feature timing	Retain event and receipt times; model configurable decision/order latency
Trade-only fixture cannot identify queue dynamics	Fill estimates are assumption-driven	Label fill results as proxy-based; keep book model interface separate
Overlapping horizons leak across folds	Inflated validation estimates	Purge by label end time plus optional embargo; boundary tests
Small/nonstationary sample	Unstable or meaningless inference	Report uncertainty and evidence tier; defer economic claims until M8
Class imbalance/calibration drift	Misleading probabilities	Persist class rates, Brier/log loss, calibration diagnostics by fold/regime
Fee/latency/fill assumptions dominate P&L	Fragile strategy result	Publish assumption grid and sensitivity; never collapse it into model quality
Public endpoint/schema changes	Broken ingestion	Version adapter/schema, capture response metadata/checksum, contract tests
Memory/disk pressure	Local run failure	Byte-bounded response streams, hard retained-evidence reservations, disk-backed incremental DQ, batch Parquet, verified spill-capable scans, explicit eager guards; measure peak RSS on any successful full L2 production before promotion
Multiple testing	False discoveries	Predeclare core hypotheses, report all tested variants, use adjusted interpretation
Missed or invalid prospective L2 session	Outcome-driven replacement or false continuity	One frozen calendar and campaign identity; atomic INSUFFICIENT_DATA; no replacement date; require observed-interval and cross-symbol coverage gates
Source/runtime/root changes across the four captures	Sessions are not one prospective campaign	Root campaign authority binds config/protocol, outcome-blind nonce, canonical output-root filesystem identity, clean source/import origin, and hashed Python/platform/dependency runtime; reject drift before network access
Held-out L2 economic access without a development authority	Selection leakage	Verify Aug 10/11 control authorities and first persist either eight child locks plus one aggregate LOCKED authority or a control-only NOT_CREATED authority; only the LOCKED branch may expose Aug 12/13 frames, and its evaluator has no fit path

Acceptance test matrix

1. Reproduction: a clean environment follows README.md, runs make reproduce-sample, and receives a run directory with provenance, validation, folds, metrics, trades, sensitivities, and reports.
2. Data: normalization is deterministic; timestamps are UTC; manifest hashes match bytes on disk; Parquet is partitioned by source/symbol/date.
3. Book: snapshot-plus-delta replay enforces side sorting, nonnegative depth, sequence continuity, and uncrossed top of book.
4. Timing: deliberately shifted future values cause leakage tests to fail; label intervals never enter feature windows or training folds.
5. Models: historical/majority, logistic or regularized linear, and tree models run on identical time folds without selecting on the final test set.
6. Execution: fees, two latency components, fill probability/queue proxy, partial fills, adverse selection, inventory cap, liquidation, turnover, and size/capacity sensitivity are represented and unit tested.
7. Reports: model table and technical report are rendered from serialized results; each carries data interval, config hash, manifests, and Git state.
8. Claims: synthetic and smoke outputs contain an explicit non-empirical banner; no unverified profitability or significance claim is emitted.

Definition of done

The objective's ten acceptance criteria have reproducible evidence for M0-M7:

1. README.md provides a clean setup and reproduction path.
2. The canonical workflow is deterministic synthetic data and needs no download.
3. Snapshot/delta sequence, stale/overlap/gap, crossed-book, and invariant tests pass.
4. Feature/label lineage, strict-before joins, censoring, and deliberate leakage tests pass.
5. Unpenalized and regularized logistic models plus a shallow tree share frozen OOT folds.
6. Execution records fees, two latency stages, queue/fill uncertainty, partial fills, adverse selection, inventory, liquidation, turnover, and size sensitivity.
7. Reports are rendered from serialized, checksum-verified outputs.
8. Run provenance records actual UTC coverage, config/input hashes, seed, runtime, Git revision, dirty state, and exact tracked/non-ignored source-tree digest.
9. Synthetic watermarks and claim checks prevent an unverified profitability claim.
10. Limitations, evidence promotion rules, and the absence of evaluable failed hypotheses are documented.

Passing M0-M7 completes the portfolio-quality system and first vertical slice. M8 remains the active empirical milestone. Its prospective calendar, hypotheses, selection rule, untouched tests, uncertainty, failure policy, and lock-before-open boundary are frozen in docs/M8_MULTIDATE_TRADE_PROTOCOL.md. The raw-only acquirer and terminal producer are implemented and adversarially tested. All eight full daily archives and their official evidence are bound by one verified raw-only manifest, without opening a CSV member during acquisition. The subsequent clean-commit economic run reached a verified INSUFFICIENT_DATA terminal at the ETHUSDT training DQ gate: no selection or held-out member access occurred, so the trade hypothesis was not evaluated and the date/policy cannot be replaced.

Book claims now depend on the separately frozen live-L2 campaign. The immutable capture rules are in docs/M8_L2_PROTOCOL.md; the exhaustive downstream rules and their exact source/semantic hashes are in docs/M8_L2_ANALYSIS_CONTRACT.md. Remaining acceptance evidence is: the four exact session terminals sharing one clean campaign authority; a durable LOCKED | NOT_CREATED development authority before held-out access; unchanged no-refit evaluation on the LOCKED branch; generated descriptive/predictive/execution artifacts or an honest INSUFFICIENT_DATA terminal; peak-RSS evidence below the local ceiling; and clean-room reproduction. The producer, recursive verifier, CLI/Make interfaces, self-contained authority snapshots, and external non-mutating report renderer are implemented and covered by offline tests. The existing public sample remains exploratory only.