Insightconsultance.ai

AI Hedge Fund (virattt) Review in 2026

8 min read

Muhammad AliMuhammad AliFounder, consultance.ai
AI Hedge Fund (virattt) Review in 2026

virattt/ai-hedge-fund is a free, MIT licensed Python project that runs a simulated hedge fund, with LLM investor agents and quant models forming views and code handling sizing and risk. It backtests and paper trades but places no real trades, and its README says it is for educational and research use only. Running it costs market data (from $20 one time, commercial license from $2,000 a month) plus model API calls.

What this covers

  • 64,530 GitHub stars and MIT licensed as of 7 November 2026; installs as the aihf package.
  • Ships backtesting and paper trading with a hash chained ledger; no live broker yet.
  • Five LLM investor personas and one quant model (post earnings drift) ship in the current engine.
  • Default data from Financial Datasets; commercial data license starts at $2,000 a month.
  • No overfitting validation gate yet; point in time data handling is in progress.
  • Best used as a research and learning harness, not a system for real capital.

virattt/ai-hedge-fund is a free, MIT licensed Python project by Virat Singh that runs a simulated hedge fund. LLM "investor" agents and quant models form the views, and plain code turns those views into positions you can backtest or paper trade. It does not place real trades, its own README says it is for educational and research purposes only, and the cost of running it is mostly market data and model API calls, not software.

We build finance AI systems for funds, family offices and deal teams, so we read this repo the way a buyer would. We looked at what ships today, what the license covers, what it costs to run, and where it stops. Everything below comes from the repository and from independent write ups we fetched on 7 November 2026.

What AI Hedge Fund is

The ai-hedge-fund repository describes itself as "a proof of concept for an AI-powered hedge fund." On 7 November 2026 it had 64,530 GitHub stars and 11,339 forks. It was created in November 2024, is written in Python, and had its latest commit on 3 November 2026, a version bump to 2.7.0. It is published on PyPI as the package aihf, which needs Python 3.11 or later.

The project's VISION.md sets out the model: the org chart of a real fund with the humans swapped for AI. Analysts produce views, a portfolio construction step turns views into target positions, a risk model applies hard limits the analysts cannot override, and a ledger records every decision. The same document is candid that "much of this is aspirational" and points to the roadmap for what is built versus planned.

If you read about this repo earlier in 2026, note that it changed shape. An April 2026 hands on review described 19 investor agents orchestrated with LangGraph. The current version is a rewrite: the roadmap says the team is "retiring the v1 CLI" and that the remaining investor personas are still being ported to the new interface.

Hands turning the pages of a tabbed diligence document under a desk lamp

AI for finance teams

Screen more deals without adding analysts

First system live in about six weeks. You own the code.

Book a discovery call

What it actually does today

The current release has two modes inside a terminal app launched with the aihf command. The roadmap also lists a machine readable CLI that returns JSON.

Backtesting

Backtest mode replays a fund over history and draws its equity curve against the mandate's benchmark. Fund definitions, called mandates, are YAML files that carry no tickers, so you choose the universe for each run. The roadmap marks a transaction cost model as shipped, charging half spread and commission on every fill. It marks point in time data correctness (no lookahead) as in progress, and a validation gate for overfitting (CPCV and probability of backtest overfitting) as planned, not built.

Paper trading

Paper mode runs a fund on real market days with fake money. Each session is appended to a hash chained ledger on your machine, so the net asset value builds into a track record instead of resetting. There is an approval step that shows the exact decision before it executes, a kill switch to halt and resume a fund, and a aihf paper tick command the roadmap describes as idempotent and safe to run from cron. A calendar aware scheduler, notifications and a heartbeat are listed as next.

The analysts

Analysts come in two kinds that share one interface and both return a conviction between -1 and +1 plus a written thesis. The roadmap marks five LLM investor personas as shipped in the new engine: Warren Buffett, Charlie Munger, Benjamin Graham, Peter Lynch and Stanley Druckenmiller. VISION.md calls these "stylized approximations" of public investing philosophies, not the people themselves. One quant model ships: post earnings announcement drift. Momentum, mean reversion, regime detection, value and quality factors, and statistical arbitrage are all listed as planned.

Four strategy templates ship in the strategy library: fundamental long/short, deep value, inflections and earnings drift. Multiple strategies can run at once and are netted into one book before fund level risk limits clamp it. Capital allocation across strategies is static for now; risk parity and dynamic allocators are on the roadmap.

David Arias, CFA, ran a $100,000 paper fund on the tool for the 9 October 2026 session and traced every number through the code. His summary is useful: investor agents form views, and plain code decides the money. He also noted that the investors read fundamentals and earnings only, with no news or price trend input.

License: what MIT allows, and what it does not cover

The repository is under the MIT License, copyright 2024 Virat Singh. MIT is permissive. You can use, modify, fork and embed the code in commercial work, including inside a firm, as long as you keep the copyright and license notice. It comes with no warranty.

Two things sit outside the code license and matter more for a fund. First, the project's own disclaimer says it is not intended for real trading, gives no investment advice, and that the creator assumes no liability for losses. The license lets you do more than that; the author is telling you what it was built for.

Second, the data has its own license. The default data provider is Financial Datasets. Its Credits and Personal plans carry a personal license. A commercial license, redistribution rights and uptime SLAs start at the Professional plan. The README's wallet based pay per call option is also described as "non-professional use." A firm using the code commercially needs to license the data separately.

What it costs to run

The software is free. Running it has three cost lines.

Cost lineWhat the sources say
SoftwareFree, MIT licensed, installs with pipx install aihf, uv tool install aihf or pip.
Market data (Financial Datasets)Credits: $20 one time, 1,000 requests, 1 year of history, personal license. Personal: $200 per month, 100,000 requests, 30+ years of core history, personal license. Professional: $2,000 per month, 2,000,000 requests, commercial license. Enterprise is custom.
Model APIOne key from a supported provider: Anthropic, OpenAI, DeepSeek, Google, xAI, Kimi or TypeSafe. Cost depends on the provider's pricing and how many analysts and tickers you run. A free local path through Ollama is listed as "next" on the roadmap, not shipped.
InfrastructureRuns locally from a terminal and stores keys, mandates and ledgers under ~/.hedge-fund/. No hosted service is required.

The data tier has a practical effect on research quality. Arias found that each investor agent needs four filed quarters, and the entry data plan covers 12 months, so Apple and Microsoft did not fit his run and he used stocks that did. An April 2026 review of the earlier version put model spend at roughly $0.15 to $0.30 per run on its configuration. That number is from the old 19 agent design, so treat it as a rough order of magnitude rather than a quote.

A shelf of neatly stacked bound reports in raking side light

Build library

See what we have already shipped

Working systems, the stack behind them, and what they cost to run.

Open the library

Honest limits: research tool versus production system

The limits are documented by the project itself, which is to its credit. Here is what a fund should know before relying on any output.

  • No live trading. The README says the system does not make any trades. The roadmap lists a live broker plugin for Interactive Brokers or Alpaca as planned and off by default. All fills are simulated.
  • Backtests are not yet guarded against overfitting. The validation gate is planned. Point in time data handling is in progress. Arias flagged the missing overfitting check in his test.
  • Narrow inputs. In the current engine the investor agents read fundamentals and earnings. Alternative data connectors are on the roadmap.
  • LLM variability. An independent review by Marketmaker.cc points out that decision quality depends on the model chosen and that models with temperature above zero can give different answers to the same inputs. That matters for reproducibility and for any audit trail.
  • Output interpretation. A user testing the earlier version reported in GitHub issue #624 that missing news data came through as "NEUTRAL" and that some trade suggestions looked stronger than the evidence. They concluded it worked best as a research and scanner tool.

None of this is a criticism of the project. It is a well documented open source research engine with tests, a clear architecture and an honest disclaimer. It is simply not built to be the system of record for real capital.

Who should use it, and who should not

Good fit: analysts and engineers learning how multi agent investment systems are structured; quant and research teams who want a clean, pluggable harness to prototype an alpha model; and investment committees who want to see how separating view generation from sizing and risk works in code. The roadmap invites contributions of new analysts, strategies, allocators and data connectors.

Poor fit: anyone who wants a system that trades real money, produces auditable investment recommendations for clients, or runs on the firm's own licensed data without engineering work. The author says it is not for real trading, and the data needs a commercial license for professional use.

Where it fits for a fund, family office or deal team

For a hedge fund or multi strategy shop, the useful part is the architecture: analysts behind one signal interface, a portfolio step that turns conviction into weights, a risk layer the analysts cannot override, and an append only ledger. Those are sound patterns. Getting from there to production means the parts the roadmap lists as unbuilt: validated backtests, point in time data, a broker connection, scheduling and monitoring, plus your own compliance controls.

For a family office running AI on its own portfolio, the repo is a good way to understand what agent based analysis can and cannot do before committing budget. The real work in a family office is rarely signal generation. It is consolidating statements, monitoring managers and answering the principal's questions from the family's own documents.

For private equity and deal teams, the gap is wider. This project analyzes listed equities from a market data API. Deal work runs on data rooms, CIMs, QoE reports and credit agreements. The multi agent pattern transfers, but the inputs, checks and outputs are different. That is the problem our deal desk systems and AI financial due diligence work is built around.

A quiet corner of a modern office at dusk with city light through the window

Implementation services

Put AI into the workflow you actually run

Diligence, close, reconciliation, reporting — in your environment.

See how it works

When an in-environment build is the better answer

Fork the repo when the goal is learning or prototyping a signal. Build in your own environment when the system has to run on your data, under your data licenses, inside your security perimeter, with outputs your compliance team can audit. At that point the open source project is a useful reference, not a foundation you can lift as is.

That second case is what consultance.ai does. We build custom finance AI that runs in the client's environment, the first system is typically live in about six weeks, and the client owns the code. If you want to see the patterns we use, our library of finance AI prompts and packs is free, and you can book a call to talk through a specific workflow.

Sources

Bring one workflow to a 30-minute audit.

Book an audit