coinback — crypto futures automated trading validation project

coinback — crypto futures automated trading validation project

Project root: [original local path withheld] Git root: this directory Implementation repository: backtest/ (original supporting file) Last experiment record: E-25 (2026-07-16) Backtest README update commit: 03098f5 (2026-08-03) Top-level README updated: 2026-08-03

This directory is the project root that keeps together the design, data validation, backtests, execution research, experiment records and externally shared material of a crypto futures automated trading system.

Looking only at backtest/ shows the implementation and experiment engine, but not the original design documents, the summary report or the real liquidation sample. This README is the overall map including that outside material.

1. Conclusion first

In one sentence, the project currently stands as follows.

A substantial validation infrastructure and experiment record have accumulated, but 0 strategies have been admitted to live trading.

The purpose of this project is not simply to make backtest returns large. It is to validate whether strategies survive real execution, costs and latency, and, including failed attempts, to build a low-deviation validation system and an auditable track record.

Currently confirmed facts:

• Decision records: all of D-001~D-035 exist
• Experiment records: all of E-01~E-25 exist
• Existing backtest Git: 25 commits, main and origin/main match, no dangling commits
• Backtest tests: 117 passed (run in the backtest directory)
• Combined effective trials: N=10,342 (per the D-032 trial hierarchy)
• Strategies admitted to live trading: 0
• API order transport and real BBO/L2 queue calibration: not yet complete

2. Overall document map

The fastest reading order is:

1. This README — overall project boundaries and the current conclusion
2. backtest/README.md — the actual code tree and how to run it
3. backtest/SESSION-HANDOFF.md — what to do next
4. backtest/docs/decisions.md — decisions that must not be reversed
5. backtest/docs/experiment-log.md — the original figures and verdicts
6. trading-system-master-plan.md and the summary DOCX — design background and external explanation

3. Actual directory structure

coinback/
├── README.md                                      # entry point for the whole project
├── trading-system-master-plan.md                  # original design, Phases 0~8
├── all-experiments-report_2026-07-16.docx         # summary experiment report
├── Binance_BTCUSD_PERP_liquidation_sample_2024-10-14/ # directly verified COIN-M sample
│   ├── README_KO.md                               # meaning, caveats and hashes of the sample
│   ├── *.original.zip / *.csv                     # original, extracted and cleaned data
│   └── normalize_liquidations.py                  # code to regenerate the cleaned version
└── backtest/                                      # runnable research system
    ├── README.md                                  # entry point for the implementation repository
    ├── SESSION-HANDOFF.md                         # session continuity document
    ├── config/                                   # preregistered settings
    ├── src/                                      # data, signal, engine, execution and research code
    ├── scripts/                                  # collection, analysis and experiment CLIs
    ├── tests/                                    # currently 117 tests
    ├── docs/                                     # decision, experiment and architecture records
    ├── data/                                     # large local data
    └── reports/                                  # ledgers, SQLite, experiment results

The parquet files and raw collections in backtest/data/ are large or regenerable local outputs, so most are excluded from Git. By contrast, the ledgers, verdicts, SQLite and experiment results in backtest/reports/ are evidence records the existing repository actually tracked, so they are preserved in the top-level Git as well. Data boundaries are managed with backtest/data/split_manifest.json and SHA-256.

4. Project data flow

flowchart LR
    A["Master plan · preregistration"] --> B["Binance public API · official archive"]
    B --> C["backtest/src/data collection · normalization · quality checks"]
    C --> D["Sealed partitions · AggTrade · liquidation sample"]
    D --> E["signals + engine + research"]
    E --> F["OHLCV · event execution F0/F1/F2 replay"]
    F --> G["reports + experiment registry"]
    G --> H["harness ledger + D/E verdict records"]
    F --> I["shadow ledger"]
    I -. "real transport not connected" .-> J["paper / testnet / live"]
    K["Summary DOCX · external review"] --> H

The key boundary is signal → order intent → fill → costs · funding · markout → ledger → verdict. OHLCV alone is not claimed to reveal real queues, partial fills or latency.

5. Record preservation audit results

Existing Git history

Before moving the existing [original local path withheld] repository into the top-level Git, the following was checked.

The existing Git commits preserve, in time order, the implementation and verdicts covering data collection, partition sealing, the final test, funding, cascade, rotation, the harness, pairs, AggTrade, the maker audit and liquidation-proxy instrumentation.

Records that were outside the existing Git

The following material was outside the boundary of the old backtest Git, but is added to tracking when moving to the whole-project root.

• The original master plan trading-system-master-plan.md
• The summary report for external sharing, the 2026-07-16 all-experiments report DOCX
• The Binance COIN-M liquidation sample originals, cleaned versions, regeneration code and Korean README
• Project guidance from the parent workspace

In other words, missing records of documents, code and small validation samples were identified and raised to the top-level boundary. Conversely, the parquet files, raw collections and large caches in backtest/data/ are intentionally not put in Git. They are outputs that must be backed up separately together with split_manifest.json, SHA-256 and the experiment documents. The ledgers, verdicts, SQLite and CSVs in backtest/reports/ are experiment evidence the existing Git tracked, so they are included in the top-level Git.

6. Current project status

Completed and verified

• BTC 4 years × 4 resolutions, spot 1h, a 13-symbol 1h universe, funding for 22 symbols collected and quality-checked
• train/val/test splits and SHA-256 sealing
• OHLCV market-order, limit-order and event execution engines
• F0 touch / F1 trade-through+queue / F2 risk-averse queue fill models
• Grid search, walk-forward, Monte Carlo, sizing A/B, DSR, cointegration tests
• Official AggTrade checksum verification and replay of selected windows
• COIN-M liquidation archive deduplication and throttle consistency checks
• 117 unit tests

Not yet admitted

• Live strategies: 0
• Real Binance order transport: not connected
• Queue calibration based on real BBO/L2: insufficient data
• Post-only rejection rates, cancel races, real latency: unverified
• Independent verdicts from forward USD-M liquidation data: waiting for VPS and collection

Per-strategy verdict figures and next priorities are kept in the backtest README (original supporting file) and backtest/docs/experiment-log.md (original supporting file). The top-level README does not duplicate the numbers, to prevent them from diverging between two documents.

7. Why it is split this way

The top-level and lower-level READMEs have different roles

The top-level README explains "what is in this project, where past work remains, and which materials are outputs outside Git". backtest/README.md explains "how to actually run the code, and which strategies, engines and data boundaries currently exist".

Merging the two would bury the overall history and the implementation details in each other. Conversely, without both, the next person might see only the top-level master plan and assume it is already implemented, or see only the lower-level code and not know why particular strategies must not be retried.

Separating the original design from the current implementation

trading-system-master-plan.md is the target state. backtest/docs/master-architecture-implementation.md and backtest/README.md are the actual implementation state. The paper and live stages in the target document must not be read as currently complete.

Distinguishing originals, cleaned versions and experiment results

The COIN-M sample folder is an evidence bundle built so the actual original ZIP and the cleaned version can be compared. It is not the full research dataset, nor a USD-M live validation result. In the same way, the numbers in backtest/reports/ must be read together with ledgers, experiment documents and data hashes.

8. Entry points for running

To check the whole project, start from the top level.

cd [original local path withheld]
Get-Content .\README.md

# implementation and test work
cd .\backtest
python -m pytest -q                 # currently 117 passed

Data collection and experiment commands, and the rule against reopening the test set, follow backtest/README.md (original supporting file) . The DOCX and sample outside backtest are explanatory and audit material, not an automated trading execution path.

9. Git boundaries and operating rules

The Git reference point is now [original local path withheld]. backtest is no longer a separate Git repository, and the existing backtest commit history is brought under the backtest/ path of the top-level repository.

• The top-level README.md, master plan, summary DOCX and COIN-M sample are managed by the top-level repository.
• Code, documents, settings and tests in backtest/ are also managed by the top-level repository.
• Large raw data in backtest/data/ are excluded by .gitignore, and split_manifest.json is tracked.
• The ledgers, verdicts, SQLite and CSVs in backtest/reports/ are experiment evidence and are tracked in Git.
• The existing Git remote backtest.git is preserved only as the source of past history. The top-level repository's remote and push target must be decided separately.
• Before committing, check git status, git diff --check and cd backtest; python -m pytest -q

.

Update rules

The following changes update the top-level README and lower-level documents in the same commit.

The top of the README keeps at least the following three dates.

1. Last code reference commit
2. Last experiment record
3. README cross-check/update date

10. Cautions when reading the whole project

• "There are many strategies" does not mean "there are many deployable strategies". Currently 0 are admitted.
• "There is a master plan for going live" does not mean "a live trading path is connected".
• "There is a liquidation sample" does not mean "all USD-M liquidations have been reconstructed".
• "There are reports and ledgers" does not mean "the data files are included in Git".
• Do not revive figures or parameters without reading backtest's detailed README and its decision and experiment documents.

As of now, this root README must answer three questions.

1. What is being built? — a crypto futures research system that builds a validated track record
2. How far has it come? — infrastructure and records substantially complete; 0 live strategies
3. Does the past work remain? — the D/E documents, existing Git commits, master plan, summary report and liquidation sample were each checked and preserved, and are now managed together within the top-level Git boundary.