How to use AI for quant trading strategies

Use AI for the research and the engineering, and let tested code do the trading. AI is good at turning a trading idea into explicit rules, writing the code and attacking the results. It is a poor judge of its own backtests, and a language model shouldn’t be deciding trades live. With Senpi you can work two ways. You can research and write strategies in the AI you already use (Claude, ChatGPT, Grok, Muse or Hermes, connected to Senpi), or you can do it all directly on Senpi. Either way, you describe the strategy in plain English and approve its rules, and Senpi’s runtime trades those rules as code on Hyperliquid, 24/7, exiting through a ratcheting stop and within the loss limits you set.

Use your own AI, or do it all on Senpi

You don’t have to switch assistants. Connect the AI you already use to Senpi and research and write strategies there, or do everything in the Senpi app. Either way, the strategy runs on Senpi’s runtime under the same rules and approvals.

Connecting takes one address, https://agents.senpi.ai/mcp, and your Senpi sign-in, with no API keys to paste. The setup steps for each app are in Can ChatGPT or Claude trade crypto?. Developers can add the same address as a custom connector in Claude Code or Codex. The open-source Senpi — AI Quant plugin brings Senpi’s research tools to Cursor and other coding agents; it is read-only and can’t place trades.

Two ways to use AI in a strategy

AI can build the strategy, or it can be part of the strategy. Start with the first: it gives you a research process you can trust before you add model complexity.

Senpi draws the line there on purpose:

• Code is reproducible. The same inputs give the same trade, so a live result can be compared with what the rules said.
• Code is auditable. The strategy is a readable package, and Senpi’s strategy templates are open source on GitHub.
• It’s cheaper to run. Senpi’s runtime supervises a strategy around the clock without spending AI credits. Credits pay for conversations, not for a strategy that is already running.

A practical workflow on Senpi

1. 01

   Choose a narrow problem

   Pick the market, the timeframe and what you’re measuring against, such as holding BTC. Hyperliquid’s perps cover about 180 crypto markets and about 120 XYZ markets for stocks, indices, commodities and FX, all trading 24/7. Decide your budget, your largest acceptable drawdown, your leverage and whether shorting is allowed. On Senpi these become the strategy’s guard rails: a daily loss limit, a drawdown halt, a cap on entries per day and a leverage limit.

2. 02

   Turn the idea into a testable hypothesis

   Describe it in plain English, in the Senpi app or in your own assistant, for example a mean-reversion rule. Senpi’s agent asks one decision at a time: what to trade, how to score it, how to exit, how much to risk. It reads the whole spec back to you before it writes any code. If a template is close to your idea, start from it instead: it deploys under your name and you change its levers.

3. 03

   Audit the data

   Hyperliquid’s public API is the primary source, and it has limits that shape any test. It serves only the latest 5,000 candles per interval: about 3.6 days of 1-minute bars, 208 days of hourly bars and 833 days of 4-hour bars. 56 of the 235 perps listed on the main book are delisted, so a test that uses today’s list alone carries survivorship bias. Timestamps are UTC. At weekends, XYZ stock perps are priced from Hyperliquid’s own order book (how to trade stocks on Hyperliquid).

4. 04

   Write the rules and price the costs

   Say when a signal forms, how orders execute, how positions are sized and how they exit. On Senpi, the runtime sizes each position as a share of the account and places orders maker-first to save fees. It exits through DSL: a protective floor first, then profit locks that only move up. Count the fees, the hourly funding and the slippage a stop pays when it fires.

5. 05

   Use AI to attack the results

   Ask the agent to try other periods, higher costs, delayed fills and nearby parameters, and to check whether a few trades carry the whole result (profit factor helps here). Keep a log of every variant: picking the best of many tries is how backtests overfit. If a template cites a replay, read it as one past window, not a result.

6. 06

   Validate, then go live small

   Build on earlier data, test on later data and keep a final window you never touch. Senpi has no paper-trading mode yet, so a strategy is tested in two steps. First, a validation run proves it runs. Then it goes live at its minimum budget, with the daily loss limit and the drawdown halt on. Every fill is public on Hyperliquid, so you can compare live fills with what the test assumed. After it’s live, ask the agent to review your closed trades for fees, exits and sizing.

Hyperliquid checks before you trust a backtest

Most strategies that fail live fail on these points, not on the idea itself.

Check the code an AI wrote

An AI writes plausible strategy code fast, and the most common bug in it doesn’t show in the results: the strategy uses information it wouldn’t have had when the order went in. Check these before you trust any number:

• Lookahead. Every signal must use only data that existed before the order. A rule computed on a bar’s close can trade only from the next bar; code that trades on the current bar’s close, high or low, or on a daily value before the day ends, is peeking.
• Survivorship. Test on the markets that existed at the time, delisted ones included, not only today’s list.
• Costs. Charge taker fees on every market order and stop, slippage, and hourly funding, as in the table above. Code that leaves them out overstates every result.
• Timestamps. Hyperliquid’s candles are stamped with the time the bar opens. Check time zones and gaps, and apply funding at the hour it was paid.
• Overfitting. Count the variants you tried, and keep a test period the AI never saw.

A worked example of lookahead. On a random walk, where no rule can have an edge, the same 20-bar trend rule “earns” +623% when it trades the bar its signal was computed on, and +8%, which is noise, when it trades the next bar:

import random

random.seed(7)
closes = [100.0]
for _ in range(2000):                      # a random walk: there is no real edge to find
    closes.append(closes[-1] * (1 + random.gauss(0, 0.01)))

def backtest(delay):
    pnl = 0.0
    for t in range(20, len(closes) - 1):
        ma = sum(closes[t - 19:t + 1]) / 20                # 20-bar average, includes the close of bar t
        signal = 1 if closes[t] > ma else -1                # long above the average, short below
        i = t + delay                                       # delay=0 earns bar t's own move: lookahead
        pnl += signal * (closes[i] / closes[i - 1] - 1)
    return pnl

print(f"trading the bar the signal was computed on: {backtest(0):+.0%}")
print(f"trading the next bar:                       {backtest(1):+.0%}")

Ask the AI to do the same to its own code: move every signal one bar later and rerun. If most of the edge disappears, it was lookahead. Then ask it for a test that fails whenever a signal reads a value from the bar it trades or later.

Prompts to start with

These work in the Senpi app, and in Claude, ChatGPT, Grok, Muse or Hermes once it’s connected to Senpi.

Turn an idea into a test

Act as a skeptical quant research partner. I want to test this idea: [your idea].
Before you build anything, give me the reason it might work, the data and timestamps
it needs, the exact entry, exit and sizing rules, the benchmark, the costs on
Hyperliquid (fees, funding and slippage) and the result that would make us reject it.
Start with the simplest version, list every variation we try, and don't call it
profitable from a backtest alone.

Start from a template

Which Senpi template is closest to this idea, and which of its levers would I change?
[your idea]

Set the guard rails before it goes live

Before this strategy goes live, set a daily loss limit and a drawdown halt on it,
and tell me what one stop-out costs as a share of the account.

Review what happened

Review my closed trades from the last 30 days. Where did fees, exits or sizing
cost me the most?

Read the market first

What does smart-money positioning look like on HYPE right now, and is it crowded?

Frequently asked questions