ALPHAC — the quant engine behind canlicapital.com

ALPHAC is a cross-asset, market-neutral research and trading system, and this is all of it: the data lake, the point-in-time reader, the backtester, the walk-forward harness, the multiple-testing machinery, the portfolio optimizer, the live broker loop, and every design document and adversarial review that produced them.

Created and maintained by Arhan Canli for Canli Capital. Development uses reviewed AI-assisted tooling, but project ownership, research decisions, methodology, claims, and publication responsibility remain with Arhan Canli. Citation metadata is provided in CITATION.cff.

It is public because the claim we actually make is not "this makes money." It is "every number we publish can be checked, including the ones that embarrass us." That claim is worthless if the code is hidden.

The honest position, stated first

We run this on paper capital. No real money has been deployed.

Evidence snapshot: 2026-09-06. Later marks must update this table through the same artifact-bound publication pipeline; this is not a real-time broker display.

Paper sleeves	4 / 14 planned — funding carry, equity momentum, managed-futures trend, PIT macro surprise
Forward record	28 daily returns from 2026-08-07 through 2026-09-06; cumulative return −3.21965%; provenance currently passes the publication gate
Forward Sharpe	Not reportable — 252 observations are required for an estimate and 756 for the project's establishment test
Drawdown	Realized 3.32408% to date, descriptive only; the current-composition model estimates 9.318% expected / 16.451% p95, neither established by live evidence
Diversification	Research-curve average pairwise correlation +0.02483 across 4 sleeves; live-forward diversification is not established
DSR policy	Mandatory to measure and publish; 0.95 is a full-union portfolio-maturity threshold, not a per-sleeve or incremental-admission gate

No forward Sharpe or expected maximum drawdown is established. The 28-return record is too short, and its provenance gate currently passes. Historical simulations, modeled risk and broker-derived paper marks remain separately labelled; none is a promise about future returns.

What is actually interesting here

The strategies are well-documented academic families — cross-sectional momentum, funding carry, time-series trend, macro-surprise rotation. We are not claiming a secret. What took the work, and what this repo is really for, is the machinery that stops us fooling ourselves:

• A union trial ledger. Deflated-Sharpe correction counts every hypothesis ever tested across every durable research ledger, deduplicated — 229 total identities: 228 retired legacy identities plus one sealed prospective identity that ended incomplete and was not admitted. Which directory a trial landed in is a filing convention; multiple-testing correction does not care. Prospective research has a staged 400-identity ceiling, with hard reviews at 320, 360 and 400. The remaining capacity buys permission to test, never permission to admit.
• Pre-registration that can be voided. Each candidate names its mechanism, data vintages, costs, variants and kill condition before the holdout is read. docs/design/PREREG_*.md. A run that deviates from its own document is not quietly accepted.
• A published kill log. Dead ideas keep their full return curves, not a scalar verdict — you cannot build a portfolio out of scalars. Recent kills: all five PIT macro-surprise series, clustered insider purchases, crypto VRP, eight AlphaMax construction variants.
• docs/retracted_claims.txt. Numbers we published and later found wrong, with the correction and the date. Two of the largest: average pairwise correlation was published as ~−0.02 when the three-sleeve book measured +0.0723 (the current four-sleeve research study measures +0.02483), and AlphaTrend was called our soundest sleeve at DSR 0.83 when its honest DSR is 0.000 — it had been graded with a variance input ~80x too small, an easier exam than its siblings.
• A signed, Bitcoin-anchored transparency chain. An append-only hash chain over the published record, so a past claim cannot be silently rewritten. The verifier and anchoring scripts are here; the log itself is operational state, published at canlicapital.com/open rather than committed.
• Tests that pin the path that runs, not the intention. Twice in this project a fix was correct, unit-tested, and never executed in production — funding was booked zero times in 44 days behind a swallowed TypeError, and a regime detector fires on 0.02% of days when it should fire on 81.8%. Both are documented. The lesson is enforced in the test suite.

Known-open defects

Kept here rather than in an issue tracker nobody reads:

1. Overlay scale defect (portfolio / strategy.py) — the realized-vol leg measured the post-overlay equity curve while the ex-ante leg used pre-overlay weights, never the same scale, so the fast leg could not bind; estimated +0.6 to +2.2pp of expected max drawdown while open. Fixed 2026-08-18 (the realized leg is measured on the unlevered book; pinned by tests/unit/test_overlay_realized_leg_scale.py) and made to survive production on 2026-09-06 (under --once the leg's history lived in the process and restarted empty every hour; each cycle now records its overlay scale and the strategy is re-seeded on boot). This entry said "Open" from 2026-08-18 to 2026-09-06, written the same day the fix landed and never updated; the transparency log carries the correction.
2. AlphaLedger's pre-registration may be void. PREREG_SLEEVE4_INVESTMENT.md pins the universe to a frozen 8,017-id allowlist; the run that produced its headline evidence (21y Sharpe 0.83, NW t +3.19) resolved the universe dynamically and used 6,880 ids. The corrected re-run is written (scripts/rerun_alphaledger_pinned.py) and has not been executed. The sleeve is plumbed but is not in the live book, and will not be until this is settled.
3. Trial-packet backfill is mostly incomplete. The public corpus contains 101 research papers and the one-to-one manifest publishes a packet for all 228 hypothesis identities, but only 2 packets currently contain every required protocol, result, lineage and correction section; 226 remain incomplete. Family-paper coverage is not exact-trial reproducibility, so the project explicitly does not claim a complete paper and evidence packet for every trial.

Reproducing

uv sync                 # Python 3.12, pinned via uv
uv run pytest           # the suite
uv run af --help        # CLI

The Parquet data lake (data/, ~26GB) and run artifacts (artifacts/) are not in git — they are re-derivable from the exchange and vendor sources the ingest scripts name. Published numbers and their inputs are mirrored as machine-readable artifacts under canlicapital.com/open.

Every tracked sleeve-publication bundle can also be checked before the Python environment is installed:

python scripts/verify_publication_clean_checkout.py --receipt publication-integrity.json

That clean-checkout check currently covers 16 incomplete preparation bundles, 402 checksum-bound files and 80 machine-validated PDF pages. It verifies archive integrity, authorship, current code/environment bindings and fail-closed publication claims. It does not regenerate returns, clear data rights, establish independent replication or claim submission.

Not investment advice

This is research software published for inspection. It is not investment advice, not an offer, and not a solicitation. Nothing here is a promise or forecast of returns. Past and simulated performance do not indicate future results. Trading involves risk of total loss. Run it against your own capital and that is entirely your decision and your risk. See LICENSE — the software is provided "as is", without warranty of any kind.

---

Engine architecture

Production-oriented mid-frequency quantitative research and paper-trading system under active development. The multi-asset core includes a Binance USDT-M perpetual path on 1h bars (signal at bar close, execution at next open), spanning the data lake through live paper operations.

Non-negotiables

• Point-in-time everywhere. Timestamps are UTC epoch-milliseconds; bars are labeled by open time and become available at ts_open + Δ. A decision at close T may only use records with available_at ≤ T. The PITDataReader is the single read path enforcing this.
• One of everything integrity-critical. One Instrument model, one calendar, one feature framework, one TransactionCostModel, one purged-splitter library, one scheduler.
• No same-bar fills. Every engine fills at the next bar's open plus modeled costs.
• Survivorship-bias-free. Backfill candidates come from historical listing/delisting records, not the live instrument list.
• Full design rationale and the adversarial review live in docs/design/.

Layout

• src/alphaforge/ — the engine (core, data, features, validation, signals, costs, backtest, portfolio, risk, execution, live, analytics)
• configs/ — layered YAML settings (base.yaml + env overrides via pydantic-settings)
• data/ — Parquet lake (gitignored, re-derivable from exchanges)
• var/ — operational SQLite state + logs (gitignored, backed up nightly)
• tests/ — unit/, property/ (hypothesis), golden/ (frozen fixtures), integration/

Development

uv sync                 # install env (Python 3.12 pinned via uv)
uv run af --help        # CLI
uv run pytest           # tests
uv run ruff check src/alphaforge tests  # governed lint-clean scope
uv run python scripts/export_lint_debt_contract.py  # disclose legacy script debt
uv run mypy             # types (strict)