Skip to content
HaiNINE30

Quant research with an agent that cannot cheat.

Hai is a research environment for systematic trading. Market data as each day knew it, since 1990. Backtests, parameter studies and paper trading. An agent that does the work, inside a harness that checks it.

For systematic traders, small funds and independent quants.

What a first backtest looks likeRecorded in Hai on November 2, 2026. The rules and figures are the run's own.

You

Build me a momentum model on the S&P 500: hold the 25 strongest stocks by their last 90 trading days, only ones scoring at least 40, rebalanced once a month. Then run it over the last ten years with trading costs.

Hai

Building it. The model trades S&P 500 members as they stood on each day, with SPY as the market trend filter and IEF and TLT holding the cash.

On the first trading day of each month it ranks the stocks that clear a 20-day volatility floor by their 90-day momentum score.

It holds up to 25 names, weighted by inverse 20-day volatility, keeps a position while the stock stays in the index and the top 100, and opens new ones only when SPY is above its 176-day average and the score is at least 40.

Running the ten-year backtest with trading costs now. I will read the figures back when it finishes.

Done. Over ten years with trading costs it compounded at 15.80% a year, growing $1,000,000 to $4,314,601, with a worst drawdown of 33.77%. Simulated, not a forecast.

20162026
+331% Growth over ten years−34% Worst drawdown3,602 TradesSimulated. Not actual trading.

Growth over the same ten years, by momentum window (trading days)

  • 60+208%
  • 90+331%yours
  • 160+438%
  • 200+278%
  • 250+255%

Then it asks what you want to change, and shows the whole row.

You describe the model. The agent builds it, runs the ten-year backtest with costs, varies one setting and shows the whole row, then reads the figures back and offers to paper trade it.

Andreas F. Clenow.
Andreas F. ClenowFounder, NINE30

Built by a practitioner.

Andreas F. Clenow wrote the books on modern quantitative trading. A best-selling author in the field, he has managed systematic hedge funds for decades and was global head of equities and commodities analytics at a large institution before that.

The catalog models are the ones from the books, so you start from something that has been in print for a decade and make it your own.

From the data to the paper trade.

Each of these is a screen in the portal and an API call your agent makes with your permissions.

  • Markets and dataUS stocks and ETFs since 1990, sixty-six futures markets, spot crypto, each day as it was known.
  • Lists and universesRankings, screens and universes you define, kept current every session.
  • Analytics and labelsIndicators, correlations and risk labels computed on the server, on every instrument.
  • ModelsA model is a short program. Write it yourself, or describe it and the agent writes it with you.
  • BacktestsDaily since 1990 with costs, every order recorded with how it was sized.
  • Parameter studiesVary a setting across a range and read the whole row, not one lucky value.
  • Paper tradingRun a model forward on a schedule with simulated capital, every decision on file.
  • Snapshots and chartsPin the data to an instant so a question keeps its answer. Mint a chart and share it.

Why a figure from Hai holds.

Your model's code runs inside Hai's harness, against the data as it stood on each day, served one day at a time. It cannot see tomorrow, and neither can the agent.

The figures the agent reports are the run's own, read back from the result, never computed in a chat. If the data for a run is not there, the run refuses before it starts.

Runs are deterministic. Every run has an id, a snapshot and its orders, and runs again to the same number.

A backtest result in Hai: the equity curve against SPY over ten years, the drawdown beneath it, and the returns in the corner.

From idea to paper trading.

  • Describe Say what it trades and the rule it follows, or write the code yourself.
  • Backtest Ten years or thirty, with costs, on data as it was.
  • Parameter study Vary one setting across a range and read the whole row.
  • Paper trade Forward on simulated capital, on a schedule. Every order recorded.
  • Live Hand the orders to your own broker. Available on Hai Pro with a compatible brokerage account.
A paper trade's orders in Hai: session, instrument, side, fill, price and commission.
A paper trade, as Hai records it: every order with its session, fill and commission.

From the agent you already use, or the portal.

  • Hai is an MCP serverConnect Claude, ChatGPT, Claude Code, Codex or Cursor, or run DeepSeek or another model in an app that supports MCP, and sign in once. Your agent gets the same data, the same tools and the same permissions as the portal, and no key is ever pasted into a chat.ClaudeChatGPTClaude CodeCodexCursor
  • In the Hai portalThe agent sits beside your lists, models, runs and charts in the browser. Nothing to install.MarketsListsModelsBacktestsStudiesPaper trading

Mix them. Start a study from your terminal and read it in the portal on your phone.

Plans and pricing

Start free. Move up when you need more room. Invited members keep their terms.

  • Free

    $0

    • 500K agent tokens a month
    • 10 backtests a month
    • 1 API key, 500K platform tokens a month
    • 1 active paper trade
    • Create models and open models shared with you
    • The best agent model on paid plans
    Start free
  • More room

    Talk to us

    • Custom allowances for agent tokens, backtests and platform tokens
    • The best model for the agent
    • Live trading through your own broker
    • Don't use my data for training
    • Support and integrations

Payment and invoices are handled by Stripe. Change or cancel any time under Usage & plans.

Hai is a research and simulation platform. Simulated results are not actual trading, and nothing here is investment advice. Built by NINE30

Start free.

10 backtests a month, 500K agent tokens, 1 paper trade, for nothing. Hai Plus and Hai Pro add room, the stronger models for the agent, and the choice to keep your research out of training.

The demo, in a few minutes.