Holdout Lab
● free · self-hosted · paper-trading · honest by design

Learn algorithmic trading without fooling yourself.

Respect the holdout.

A self-hosted lab with the whole path in one app — a hands-on course, an AI strategy studio, a nine-stage anti-overfit gauntlet, and paper trading — so you find out what actually holds up before a dollar is at risk.

Runs on your machine · your data · your broker · no real money at stake

Everyone getting into algo trading hits the same wall

Flattering backtests

Every tool draws a gorgeous equity curve. None of them warn you it's overfit.

Nowhere safe to learn

The tutorials stop at a backtest. The very next step is real money.

Hype over honesty

Signal-sellers and curve-fit gurus everywhere. The truth, nowhere.

Most backtesters tell you what you want to hear.

Holdout Lab is built to tell you the truth — and that's exactly what a newcomer needs to learn.

Typical backtester · in-sample

“+312% · Sharpe 2.8 · looks amazing!”

Holdout Lab verdict · out-of-sample

“Overfit. Out-of-sample edge ≈ 0. Don’t trade this.”

every idea runs the gauntlet:rule-test vs randomwalk-forwarddeflated Sharpe2× cost stress

The whole lab, tab by tab

One app. Every step from first lesson to live-grade discipline.

Holdout Lab is a single self-hosted app. Here's what each tab does — honestly.

01

Learn & build

Start with a course that runs real backtests, then build with AI, your own data, and real charts.

Learn tab

A course that runs on the real engine, not canned demos.

Four modules take you from foundations to the overfitting trap to the gauntlet — and every lesson has you doing it, not reading about it.

  • Hands-on exercises run real backtests from inside the lesson
  • A practice library of curated strategy templates to run, tweak, and break
  • Graduate straight into your own backtests and paper sessions
The Learn tab →
Studio tab

Describe a strategy in plain English. Get runnable code — safely.

The Studio turns a sentence into a working NautilusTrader strategy with your own Anthropic key — then validates and sandboxes it before it touches a single bar.

  • Bring-your-own LLM key — no cloud service of ours in the loop
  • AST allow-list validation plus an isolated, resource-limited sandbox
  • 19 built-in strategies too — including the ones that honestly tested negative
The Studio →
Data tab

Bring your own keys. Data lands locally and never leaves.

Free daily equities, free 1-minute crypto, and pay-per-use intraday data — every pull lands in a local Parquet catalog on your machine.

  • Tiingo daily equities (free key) · Binance 1-minute crypto (no key at all)
  • Databento intraday shows a cost estimate you confirm before it spends anything
  • Backtests read only from the local catalog — no vendor calls mid-run
Data ingestion →
Charts tab

Study the same bars the engine trades.

Saved chart workspaces with drawings and indicators — and any strategy's entries and exits overlaid on the candles it actually saw.

  • Saved workspaces with drawings and configurable indicators
  • Strategy overlay marks every entry and exit on the bars
  • Identical local data in charts, backtests, and live sessions
Charts →
02

Prove it

Backtest every idea, then let the honest scoreboard tell you which ones actually earned trust.

Backtests tab

Every run kind, one honest report.

Full metrics, equity vs buy-and-hold, drawdown, and a Monte Carlo band — with no look-ahead by construction: closed bars, next-bar fills, modeled costs.

  • Sharpe, Sortino, max drawdown, round-trip stats, buy-and-hold excess
  • Fills happen on the next bar with modeled costs and slippage
  • Scales up: cross-sectional ranking, cointegrated pairs, portfolio search, multi-strategy books, crisis stress
Backtesting →
Research tab

The honest scoreboard: 0–100 readiness, and the gates that earn the star.

Every strategy × symbol × timeframe you research is scored on one board — with the four hard gates it must clear shown right next to the number.

  • Four hard gates: rule test, held-out Sharpe, walk-forward OOS, deflated Sharpe
  • Buy-and-hold excess always shown — beating cash isn't the bar
  • One click from a validated row to a live paper session
The research ladder →
03

Trade it

Forward-test with zero money at risk — on rails built like real money is on the line.

Paper tab

Zero-setup forward testing — watch your edge live.

The exact strategy object you backtested runs against live market data with no money at stake — the honest dress rehearsal most tools skip.

  • Crypto sandbox: live Kraken market data, local fills, nothing to configure
  • Interactive Brokers paper trading via a dockerized gateway the app manages
  • Multi-sleeve book sessions with telemetry, an event feed, and persistent history
Paper trading →
Live tab

Built like real money is on the line — because one day it is.

The live path exists so the discipline is real: every order must clear a pre-trade risk engine before the broker ever sees it. The free edition ships research and paper only — live trading stays edition-gated.

  • Pre-trade risk engine denies limit breaches before they reach the broker
  • Audit-logged approve-to-live gate, kill switch, reconciliation on reconnect
  • Telegram alerts and a gateway health monitor watch the infrastructure
Live trading & safety →

Every idea runs the gauntlet

Nine ways a backtest can fool you.

Every stage exists because of a specific way people lose money to their own numbers.

Re-optimized, then traded forward — never peeking.

Walk-forward optimization re-tunes the strategy on each past window and trades the next one out-of-sample, stitching the honest segments into one curve. A parameter-stability check flags strategies that are just fitting noise.

How validation works →

Know the range, not just the line.

A single equity curve is one roll of the dice. Holdout Lab runs a Monte Carlo over thousands of resampled paths, so you see the median and the 5–95% band — the realistic spread, not a cherry-picked best case.

Metrics & robustness →

How you run it

Your machine. Your keys. One command.

Self-hosted, actually

Download the bundle, run two commands, open your browser. Single-user auth, nothing hosted by us, no cloud account.

$ ./setup.sh
$ docker compose up -d

Install →

Bring your own keys

Tiingo, Databento, and Anthropic keys live in your local settings — with a Test button for each. They never touch our servers, because we don't have any.

Configuration →

One engine end to end

The code that backtests is the code that trades: the same NautilusTrader strategy object runs in backtest, paper, and live. No reimplementation gap.

Architecture →

Questions

Frequently asked

Do I need to know how to code?

No. Describe a strategy in plain English and the Studio writes it for you, or start from a validated template. If you want full control, you can write Python — but you never have to.

Are my API keys and trading data private?

Yes. Holdout Lab is self-hosted — it runs on your own machine. Your keys, strategies, and trades never leave your computer and are never sent to us.

What markets and brokers does it support?

US stocks & ETFs and crypto. Paper trading runs against Interactive Brokers (paper account) or the built-in zero-setup crypto simulator. Data comes from Tiingo (daily equities), Binance (1-minute crypto), and Databento (intraday equities).

Can it trade real money?

The engine has a complete live-trading path — pre-trade risk limits, an audit-logged approve-to-live gate, a kill switch — but the free edition ships research and paper trading only. Live trading stays edition-gated.

Is it safe to run AI-generated strategy code?

It is sandboxed — no file, network, or system access — statically validated, and run in an isolated process. Nothing can ever go live until it clears the gauntlet and you explicitly approve it.

Do I need a server or the cloud?

No. It runs locally on your own machine (via Docker). Nothing is hosted by us, and there is no cloud account to manage.

What do I need to run it?

A Mac or Linux machine with Docker and about fifteen minutes. One command brings the whole app up locally.

What does the data cost?

Tiingo daily US equities and Binance 1-minute crypto are free (Tiingo needs a free key; Binance needs none). Databento intraday data is pay-per-use — the app shows a cost estimate you must confirm before anything is billed.

What exactly is “the gauntlet”?

The anti-overfit validation pipeline — nine stages, including a rule test against random-entry twins, optimization judged on held-out data, walk-forward, Monte Carlo, deflated Sharpe, a 2× cost stress, and a buy-and-hold benchmark. Only strategies that survive every stage are marked validated.

Is the backtest performance real?

Backtests are simulations on historical data — simulated and past performance never guarantee future results. The whole point of Holdout Lab is honest validation, not flattering numbers.

What does it cost?

The research + paper-trading edition is free — download it and run it on your own machine. A paid Pro edition with live-trading connectivity is planned; leave your email on this page to hear when it opens and what it costs.

Run it today — free

The research + paper-trading edition is free and self-hosted. Download the bundle and you're backtesting in minutes.

Download freeInstall guide →

Want live trading when it lands?

A paid Pro edition with live-trading connectivity is planned. Leave your email to hear when it opens — and what it costs.

By joining, you agree to our Privacy Policy.