Vibe trading
Vibe trading, with a backtest before it trades
Vibe trading means describing a trading strategy in plain words and letting an AI build it, the way vibe coding builds software from a prompt. Keel lets you vibe trade with a check in the loop: the AI builds the strategy from blocks you can read, and you backtest it on real Hyperliquid data and paper trade it before any money is at risk.
By Keel Research Team
Updated October 18, 2026Example session
Claude + Keel- YouBuild me a MACD strategy for the big coins on Hyperliquid. Go long when MACD 12/26/9 turns up, but only when the coin is above its long-run average. Get out when MACD turns down. Then backtest it.
- ClaudeKeel's library has a strategy built on exactly that, so I'll start from it: MACD Crossover Strategy (Crypto). On four-hour bars it buys a coin when MACD crosses above its signal line while price is above a long moving average, and sells when MACD crosses back down. Capital is split equally, with a cap on any one coin. Published backtest, Aug 2024 to Sep 2026, after fees, slippage and funding: Sharpe 1.87, +408.8% total return, a worst drawdown of 33.3% and 2,746 trades.
- YouFeels like it trades a lot. Run it on daily bars instead.
- ClaudeThe library tested that setting on the same coins and dates: Sharpe 0.48, +25.8% total return, a worst drawdown of 46.9% and 248 trades. The Sharpe falls from 1.87 to 0.48. Fewer trades didn't help here: the slower bars got in and out too late. Want me to keep both versions in your account so you can compare them?
What vibe trading is
Where the term comes from
Vibe coding is building software by telling an AI what you want and accepting what it writes, without reading every line. The phrase spread in February 2025 and was the Collins English Dictionary word of the year for 2025. Vibe trading applies the same idea to trading: you say what the strategy should do, and an AI turns that into something that can run.
The appeal is speed. An idea that used to take a weekend of code takes a few minutes of conversation. The catch is the part vibe coding skips. With an app, you find out quickly when the code is wrong because the button doesn't work. With a trading strategy, a wrong rule can look fine for weeks and then lose money, so the check has to come before it trades, not after.
Source for the term's history, read October 2026: https://en.wikipedia.org/wiki/Vibe_coding
A worked session
One idea, from a prompt to a backtest
The session at the top of this page is shortened and built on a strategy from Keel's public library. Every number in it is the library's published result, which is what Keel's tools return when an assistant reads a library entry.
The second request is the kind of change people make on a hunch. It sounds harmless, and the backtest shows what it actually does before you find out with money.
What the AI produced
The strategy, as blocks you can check
This is the strategy from the session, the way Keel shows it. Reading it takes a minute, and it is the step that turns a vibe into something you understand.
From the top: prices come in for the most traded perps. One branch checks that MACD is above its signal line, another that price is above its long moving average, and both must agree before a coin is bought. A separate branch marks the exit when MACD drops back below. A trade manager holds each position between those two moments, and the last blocks size the book equally across coins.
Configuration
Globalsconfig
target_timeframe4h
UniverseassetsTop by Volume · HL Perps · 15 assets15 resolved
Executionorders
rebalanceOn Change
Factories
trend_vs_200factory
Pipeline7 steps
PriceDataLoaderData Loader
ohlcvstore
Parallel2 branches
entries
Parallel2 branches
const
ConstantForecastForecast Mapper
value1
setup
Parallel2 branches
macd_up
MACDIndicator
fast_period12slow_period26signal_period9
AboveThresholdFilterSignal Transform
threshold0
uptrend
trend_vs_200call
AboveThresholdFilterSignal Transform
threshold0
MaskAndSignal Composer
ApplyMaskSignal Composer
score_signalconstfilter_signalsetup
macd_entrystore
exits
Parallel2 branches
const
ConstantForecastForecast Mapper
value1
cross_down
MACDIndicator
fast_period12slow_period26signal_period9
BelowThresholdFilterSignal Transform
threshold0
ApplyMaskSignal Composer
score_signalconstfilter_signalcross_down
macd_exitstore
TradeManagerPosition Manager
entriesmacd_entrypricesohlcv
Parallel1 branches
macd_cross_down
macd_exitload
ExitPosition Manager
ExposurePosition Manager
EqualWeightSizerPosition Sizer
target_leverage1max_weight0.2
Globals(target_timeframe="4h")
Universe(mode="top_volume", top_n=15, market="perp", resolved=["BTC", "DOGE", "ETH", "GRAM", "HYPE", "LINK", "LIT", "PUMP", "SOL", "SUI", "VVV", "XPL", "XRP", "ZEC", "kPEPE"], resolved_at="2026-07-06T20:36:37.412320+00:00")
Execution(rebalance="on_change")
trend_vs_200 = Pipeline([
{
"px": [EWMA(window=2)],
"trend": [EWMA(window=200)],
},
Crossover(fast_key="px", slow_key="trend"),
], name="trend_vs_200")
Pipeline([
PriceDataLoader(),
Store("ohlcv"),
{
"entries": [
Pipeline([
{
"const": [ConstantForecast(value=1.0)],
"setup": [
Pipeline([
{
"macd_up": [MACD(fast_period=12, slow_period=26, signal_period=9), AboveThresholdFilter(threshold=0.0)],
"uptrend": [trend_vs_200, AboveThresholdFilter(threshold=0.0)],
},
MaskAnd(),
], name="setup"),
],
},
ApplyMask(score_signal="const", filter_signal="setup"),
Store("macd_entry"),
], name="entries"),
],
"exits": [
Pipeline([
{
"const": [ConstantForecast(value=1.0)],
"cross_down": [MACD(fast_period=12, slow_period=26, signal_period=9), BelowThresholdFilter(threshold=0.0)],
},
ApplyMask(score_signal="const", filter_signal="cross_down"),
Store("macd_exit"),
], name="exits"),
],
},
TradeManager(entries="macd_entry", prices="ohlcv"),
{
"macd_cross_down": [Load("macd_exit"), Exit()],
},
Exposure(),
EqualWeightSizer(target_leverage=1.0, max_weight=0.2),
], name="macd_crossover_crypto")
Sharpe
1.87
Total return
+408.8%
Worst drawdown
−33.3%
Trades
2,746
Published backtest on the library's default settings, after trading fees, slippage and funding. A backtest shows how the rules behaved in the past, not what they will do next.
See every setting tested for this strategyWhere it goes wrong
How vibe trading loses money, and how a backtest catches it
Most vibe-traded strategies that fail do so for a handful of reasons, and none of them show up by reading the chat. They show up in a test with real costs.
Nobody ran it
A chatbot asked for a backtest without tools writes code or describes results it never computed. A table of returns in a reply is not a result. On Keel the numbers come from the engine, and you can open every run.
Costs left out
Fees and slippage on every trade, and funding on perpetual futures. A strategy that trades often can look great without them and lose money with them. Every Keel backtest charges all three.
A rule that peeks ahead
Hand-written backtests often use a price the strategy couldn't have known yet. The curve looks perfect and means nothing. On Keel the AI assembles existing blocks instead of writing its own indexing code, so there is no hand-rolled loop to slip.
One lucky setting
Try enough versions and one will look great by chance. Change a setting near the winner and see if the result holds. If it collapses, the edge was noise.
The live bot is different code
A script written after the test drifts from the thing you tested. On Keel the strategy you backtest is the one that paper trades and runs live.
Three ways in
Pick where you want to talk to the AI
All three build the same kind of strategy from the same blocks, test it on the same engine and save it to the same Keel account, so you can start in one and carry on in another.
Keel's own AI chat
Sign up and describe the strategy in the Keel app. The AI builds it from Keel's 251 blocks and runs the backtest, and you see the canvas and the result side by side. Nothing to set up.
Claude or ChatGPT over MCP
If you already think out loud in Claude or ChatGPT, add Keel as a connector. The assistant gets 29 Keel tools to find blocks, save strategies and run backtests, and hands you a link to the result in the Keel app.
Claude Code in your terminal
One command adds Keel to Claude Code. It suits a list of experiments you want worked through, with the strategy source kept in your repo next to your notes.
FAQ
Questions about vibe trading
What is vibe trading?
Vibe trading is building a trading strategy by describing it to an AI in plain words and letting the AI write it, borrowed from vibe coding. It works best with a check in the loop: read what the AI built and backtest it with real costs before it trades.
Is vibe trading safe?
Only as safe as the testing behind it. A strategy you never tested is a guess, however confident the AI sounds. Backtest it with fees, slippage and funding, paper trade it on live prices, then start small. On Keel your money stays in your own Hyperliquid account, where Keel can place trades but can never withdraw.
Do I need to code?
No. You describe the strategy and the AI builds it from blocks you can read on a canvas. There is a code view if you want it. If you do code, Claude Code can build on Keel from your terminal.
Which AI should I use?
Whichever you already like talking to. Keel's own chat needs no setup, Claude and ChatGPT connect as connectors, and Claude Code works from a terminal. They all use the same blocks and the same backtest engine.
Can I run it live?
Yes, once you are happy with the backtest and the paper run. You connect the wallet you use on Hyperliquid and sign once, and the strategy runs on its schedule from your own account. The AI never places trades itself. Hyperliquid's crypto perpetuals are supported today. Stock and index perps are HIP-3 markets, and Keel doesn't support them yet. Support is coming soon.
What does it cost?
Starting is free. The free plan lets you build and backtest strategies, paper trade one strategy and run one live, with a 5 basis point fee on live trades. Paid plans start at $29 a month and lower that fee. The pricing page has the full limits.
Describe it, then prove it
Build a strategy from a sentence, read the blocks, and see a real backtest before any money moves.