AlphaLattice

English | 简体中文

Ask for a dated research portfolio and inspect its positions, source evidence and recorded review in one workspace.

Example run on synthetic data: a request to inspect stored results, their activity and the dated positions, shown as an edited sequence of cropped views.

Tell your Claude Code or Codex agent what you want to investigate. It uses AlphaLattice to run quantitative studies, check source evidence and bring the results to a research committee. Follow the work and its results in your coding app's built-in browser.

Every numerical input is bounded by the date's information cutoff, and verification runs check for look-ahead. The committee can decline to endorse a result. AlphaLattice runs locally and is open source.

0.2.3, now

• Network-permission questions have stable codes beside their English and Chinese words, with the question's scope and reply kept in the activity record.
• A Goal's time profile shows recorded requests, Task stages and gaps. Missing observations stay unobserved; elapsed time without an attribution is not called agent time.
• On Claude Code, the first start asks you once to run the offered command in your own terminal. Codex detaches the Host on its own. This startup step is separate from the network-permission question.

See the 0.2.3 release notes for the message catalog and the limits of these diagnostics. The earlier measured workloads remain identified by their release below.

Use Claude Code or Codex as the lead. To see what a run used, open a session's Participants in the Workbench (Team, then Sessions): it shows the models and tokens each member used.

Give your agent this repository's link and one intent:

Use https://github.com/xinglin-li/alphalattice to build me a reviewed Trend Rebound strategy from public data and show its dated research positions; open the Workbench in your built-in browser so I can follow it.

Two ways to use AlphaLattice

1. First use, in a new folder. Send your Claude Code or Codex agent the intent above, or Positions for <date>, reviewed. The agent prepares data and the strategy, produces that date's research positions and follows them through review, committee discussion and a report.

   Expect a pause for your answer. The agent stops once to ask whether AlphaLattice may keep this workspace's network on for daily updates, fetching market data and SEC filings within the default budget (3 per issuer, 600 documents and 3 GB overall). The work waits until you answer yes or no. Your yes stays in force until you tell it to stop; each network use is disclosed and recorded in activity, including automated updates.

   At the first-use report, the agent also asks whether daily updates should convene the investment committee. Decide it in the conversation, then or at any later time, and change it the same way, for example "Stop convening the committee on daily updates." Until you say yes, later updates get the portfolio manager's view on the reviewed positions. The first use does not wait for this answer.

2. Daily update, in a new session in the same folder. Say Positions for <next session>, reviewed. The agent uses the installed strategy and your standing network yes to update data and produce that session's reviewed research positions, following your daily committee setting. It works in your first use's workspace, so a later date never starts a second first use, and it asks you only about decisions that are yours.

See an example run

Example run on synthetic data: Home brings the positions date and the result's provenance together.

Example run on synthetic data: a saved risk study shows its input window and diagnostics.

Example run on synthetic data: the activity record keeps the work behind the result.

Example run on synthetic data at a browser width of 900 pixels.

Before you start

• Environment: Windows is the supported platform. Setup uses Python 3.12, uv and locked dependencies, with Claude Code or Codex. On Linux and macOS, CI checks installation and an offline first use; other use there is not guaranteed. On macOS, install the OpenMP runtime:

  brew install libomp
  

• Time: In native Windows observations for 0.2.2, with Claude Code and Haiku, first use took 121.8 minutes and the first daily took 43.6 minutes. Each component used 472 supporting listings. Recorded Host work accounted for 50.75 minutes and 19.65 minutes, respectively. The daily exceeded the 20-minute budget. These are measured workloads from 0.2.2, not runtime guarantees or a new performance measurement for 0.2.3.

• Disk and memory: Those 0.2.2 runs retained an 8.70 GB first-use workspace and 8.88 GB after the first daily. The daily grew the workspace by 180 MB; its separate 4.27 GB first backup was a one-time fill. Sampled Host private-memory peaks were 9.32 GB and 9.25 GB, both above the 8 GB budget. MB and GB use decimal units. Allow headroom for your operating system and coding app; these observations are not minimum requirements. The release notes give the workload, measurement limits, upgrade results and outstanding budgets.

• Package downloads: Locked retrieval packages install under the first use's network consent, including on a machine with an empty package cache. That consent covers installation without a separate download question.

• Agent costs: Your Claude Code or Codex access and usage costs are separate from AlphaLattice; check your provider's terms.

See Getting started and the setup procedure for installation, downloads and the permissions each step needs.

Already using AlphaLattice? Follow Upgrading before installing a later 0.2.x patch in your existing workspace.

Why AlphaLattice

• A complete research workflow. Data preparation, quantitative studies, portfolio construction, Evidence review and committee discussion lead to a delivery report in one workspace.
• Evidence you can inspect. Read source passages, information cutoffs and review coverage for the reviewed portfolio.
• Disagreement you can read. Objections, limits and the committee's final ruling remain in the record and report.
• An agent that learns from your corrections. When a goal completes, the agent offers to record what the work taught it. A correction you give is kept in your own words, with why it matters and how to apply it. These notes live in your own layer, .alphalattice/user/, which releases never overwrite. The next session reads them before it starts.
• Factors and models you can extend. Add factors and Alpha models through their registries and checked extension paths.

A worked example

A completed run asked for a reviewed Trend Rebound strategy from public data, with research positions for 2026-09-11.

1. Studies. The agent prepared public-source data and ran the strategy's model and Risk studies.
2. Dated positions. It built the portfolio and delivered research positions for the requested date, with their Risk assessment.
3. Evidence review. Evidence Analysts read the prepared source material available at the information cutoff. The chief risk officer (CRO) assessed their findings.
4. Committee discussion. Alpha, Risk, CRO and the lead discussed the dated result, keeping objections and the lead's rulings in the record.
5. Report. The delivery brought together positions, Risk, review coverage, citations and the members' recorded views.

The report kept the reservations alongside the results: the review covered the filings accepted by its source checks and available at its cutoff, and members named limits in predictive support, liquidity and capacity evidence.

Historical research views

A book is a recorded research portfolio with its holdings and results. For an installed strategy, it records the strategy's historical replay.

Historical research performance for Trend Rebound Book, 2023-10-02 to 2026-09-08, from a replay of the strategy.

Sector exposure and historical replay weights for 2026-09-08.

Rebound Return Book and Trend Rebound Book are installed strategies: ready-to-use strategy definitions that can be activated to produce dated positions. They are separate from exploratory research study portfolios. Research flows explains both routes.

Read each result's information cutoff, coverage and limits on its use. For inputs built from a current membership snapshot, earlier membership is backfilled from the initial set of securities; later additions and removals follow recorded observations. This is not a complete reconstruction of past index membership. Read the reported survivorship limits, sector backfill and historical source availability alongside the performance.

Follow the work

An earlier run's recorded Evidence review assignments, answers and acceptances.

• Workbench. Use Goals for the timeline, conversation and results, History for saved studies and reviews, and Team for sessions and usage. Review pages show which checks passed, what remains missing and how the result may be used.
• First-use permission. Your sentence opens a goal and gives the agent permission for specific steps for 24 hours from the goal's opening: prepare data, handle data issues within the declared rules, open the preparation's network, set up Evidence within the default budget and activate the strategy's completed book. It reports each act.
• When permission ends. The first-use delegation ends when the goal's submission is accepted, the goal is abandoned or its 24 hours expire. Network access opened only under that delegation ends with it. The separate standing permission is described in Two ways to use AlphaLattice.
• Your controls. Stop the strategy through your agent or on Portfolio. Stopping a goal and cancelling a Task are separate controls. The guide lists what only you decide.
• Host boundaries. The agent installs AlphaLattice and continues in the same session. Specialist tool limits apply through their instructions; checkout sessions also load the host's project configuration. Start one with cd <checkout>; claude in Claude Code or codex -C <checkout> in Codex, and follow the agent guide for the host's setup and checks.
• Languages. The Workbench offers English and Chinese. CLI detail uses --lang zh, for example alphalattice workspace show --lang zh, with English fallback.

Guides and care

Guide	What it explains
Getting started	Prepare for installation and first use.
CLI	Read answers and continue saved requests.
Research flows	Follow studies, portfolios and review.
Agents	Follow goals, sessions and specialist work.
Extending	Develop and inspect factors and models.
Data and care	Update, inspect, back up and restore a workspace.
Privacy	Read what usage records contain and disable their reader.
Downloads and data terms	Read the terms for models and data sources.

The guide index links the other reference pages.

Use backup and restore into a new directory to preserve the saved records and their integrity checks. Privacy explains recorded usage and its switch.

This release and where it is going

0.2.2, now. 0.2.0 brought a one-sentence path to a dated, reviewed research result: data, strategy, research positions, SEC evidence, committee review and a report. 0.2.1 made ordinary daily updates write only the day's changes. 0.2.2 removes repeated and unused code, registers every extensible kind in one registry, and checks each sealed input once when the day's positions are read. The release notes list every change and the measured limits.

Each minor version's scope is fixed in advance and released only after a run proves it; patches in between make the current version smoother. Each version builds on methods already proven elsewhere, adapted to AlphaLattice's own rules: dated inputs, sealed results and your authority.

The next 0.2.x patches, in order. Each holds one theme, and each publishes its measured time, disk and memory against a stated budget:

• 0.2.3: questions and messages by code. Each question only you can answer, and each message, carries a stable code beside its words.
• 0.2.4: a smoother daily update. A shorter wrap-up for the agent, a measured time for each stage, and a background service that stays up between sessions.
• 0.2.5: a measured baseline. A first use and a daily, replayed without a model, report each stage's time and the program's peak memory in minutes; later speed and memory changes are measured against it.
• 0.2.6: memory within budget. Bring first use and daily updates within 8 GB.
• 0.2.7: realistic replay. Replay real-world events (corporate actions, late reports, provider outages) through the real data client and its cache.
• 0.2.8: data checks. Check the health of the data every day, in the program and the Workbench.
• 0.2.9: nothing stored twice. A storage data model; upgrades reuse unchanged data.
• 0.2.10: caching and checking, done once. Hash once, verify at trust boundaries and recompute only what changed, borrowing from Git's index and Bazel's action cache.
• 0.2.11: less waiting on the program. The program prepares likely next steps while the agent is thinking.
• 0.2.12: history you can query by date. Ask what was known, tried and decided on a past date, with nothing learned after it.
• 0.2.13: an evaluation flywheel. One runner and fixed gates check every patch from the product's own records.
• 0.2.14: a first result in ten minutes. A short trial on a small sample, labelled as a trial.
• 0.2.15: ground for daily supervision. A store for the forward record of positions.

Later minor versions:

• 0.3 Daily supervision: "watch my strategy; tell me only what matters."
  • Factor monitors in Alphalens' terms: IC, decay and turnover, on top of the daily data-health checks.
  • Drift is flagged by CUSUM and EWMA control charts, so quiet days stay quiet.
  • At most one digest a day.
  • A forward record of positions as published, kept apart from the backtest on GIPS principles.
  • Agents' calls are scored by hit rate and the Brier score.
  • The daily run is scheduled by the operating system's own scheduler.
• 0.4 Agent research: "search for a better model, honestly."
  • Overfitting guards come before any automated search: every result reports its trial count, with the deflated Sharpe ratio and the probability of backtest overfitting, a locked holdout and purged, time-ordered validation.
  • Then search with Optuna, models in scikit-learn's estimator form, and Features named after the 101 Formulaic Alphas.
• 0.5 Your own strategy: "use my tickers and schedule; add my risk model or data source." Extensions join by registry entry. Risk methods take the sklearn.covariance form, and data sources an Airbyte-style connector contract: spec, check, discover, read.
• Later, proposed:
  • macro data as it was known at the time (ALFRED vintages), company fundamentals as filed with the SEC, and consented news;
  • survivorship-free research on a CRSP-style data model;
  • several strategies as one portfolio, by risk budgeting and hierarchical risk parity;
  • committee adjustments within your bounds, in Black-Litterman form, built only if the forward record shows the committee adds value;
  • a stable 1.0 that keeps every agent-facing surface working;
  • intraday schedules.

Previews of the next minor version land on the next branch first. Tell us what you need in an issue: requests from people using AlphaLattice decide what comes after 0.5.

Tell us where you got stuck

Open an issue with the command you were running, what you expected and the answer you received. Redact private data and credentials. For proposed code changes, follow CONTRIBUTING: open an issue first for nontrivial work, and sign the contributor agreement on your first pull request. For a security vulnerability, report privately as described in SECURITY.

About

Documentation updated: 2026-10-17.

The design principles explain how the agent, the software and the person share the work.

Created and maintained by Xinglin Li: LinkedIn or [email protected]. See governance for project decisions and contribution credit, and contributing for the agreement and workflow.

Verify a source checkout

After uv sync --locked --all-extras, build Local Web with uv run python scripts/build_local_web_ui.py --product before running the public checks. Tests check that build at session start, rebuilding once if it is missing or stale; a failed build fails the run.

Development history

AlphaLattice has been developed privately since 2026-08-06. Public commits carry the shipped changes exported from that development; releases publish the resulting versions.

Disclaimer

AlphaLattice produces research results, not orders, and places no trades. Historical research is not investment advice or a promise of future returns. A completed review records findings and limits; it is not a recommendation to trade or a guarantee of scientific validity.

The software is licensed under Apache-2.0; NOTICE records third-party material. Acquired data and downloaded model packs retain their own terms. The software licence grants no data licence. Agent access and usage costs follow your provider's terms.