Case study · July 2026

We asked Meta's AI agent for its best trading strategy.
Then we graded it.

Manus produced a strategy that reads like a hedge fund wrote it — multi-timeframe trend filters, regime detection, volatility-scaled stops, disciplined sizing. Six years of out-of-sample backtesting later: F. The reason it failed is the most useful lesson in AI-era trading.

The experiment

We prompted Manus (manus.im, now part of Meta) for its best fully mechanical BTC/ETH perpetuals strategy — exact indicator parameters, computable from price data alone, no discretion, optimized for profitability after 0.05% taker fees. We deliberately asked for its best shot and ported the result faithfully, rule for rule.

What it wrote back was genuinely impressive on paper: 200-EMA trend filters on both the 1-hour and 4-hour timeframes, an ADX(14) > 25 regime gate to avoid chop, RSI(14) oversold/overbought rejection entries, stops at 2× ATR and targets at 3× ATR, a regime exit when trend strength dies, and 1% fixed-fractional position sizing. It even specified its own conservative fee and slippage assumptions. This is textbook discipline — the kind of spec that passes a code review at a prop shop.

How we graded it

Identical pipeline to every strategy on Tessen: six years of hourly data, a strict 60/40 train/out-of-sample split, and five hard gates. We used the engine's no-lookahead conventions throughout — the 4h filter only sees 4h bars that had actually closed, entries fill at the next bar's open exactly as the spec demands, and when a bar touches both stop and target, the stop fills first (conservative). We graded at the strategy's own stated costs (taker fees + slippage, ~20bp round trip, since it specifies market orders) and re-ran at our most generous 4bp maker terms as a sensitivity check. Funding costs were not modeled — including them would only make the result worse.

The numbers

GradeF — at its own stated costs AND at optimistic maker terms
Data6 years (2020–2026), BTC + ETH perpetuals, 1h primary / 4h trend filter
Trades154 total; 67 out-of-sample (60/40 train/test split by time)
OOS win rate37.3% — vs the 44.4% breakeven its realized risk:reward requires
Realized R:R1.25 (designed for 1.5 — intrabar stop fills ate the difference)
OOS gross expectancy−8.9bp per trade, before fees
OOS net expectancy−28.9bp (its own 20bp taker+slippage model) · −12.9bp (4bp maker terms)
Exit mix86 stop-losses · 36 take-profits · 32 ADX regime exits

The five gates

Out-of-sample expectancy positive
FAIL

−28.9bp net per trade at the strategy's own cost model; −8.9bp gross, before any fees at all.

Clears the cost hurdle
FAIL

Gross OOS expectancy is negative — there is nothing for fees to eat. Hurdle was 30bp at its own taker costs (and it fails at our most generous 6bp maker hurdle too).

Robust across assets
FAIL

ETH deeply negative (−41bp/trade OOS), BTC marginal. 0 of 2 assets positive net at its own costs.

Survivable drawdown
PASS

At responsible 1% fixed-fractional sizing, median max drawdown ~10%. The risk management genuinely works.

Not overfit
PASS

Train and test agree — both negative. Manus did not curve-fit history. It wrote an honest strategy with no edge, which is a different failure than overfitting.

Why this matters

The strategy passed the overfitting gate. Read that again: Manus did not do the thing AI is usually accused of. It didn't torture the data until the backtest looked good. Train and test agree almost perfectly — both negative. It wrote a clean, honest, professionally structured strategy with no edge. Its 37% win rate needs 44% to break even at the risk:reward it actually realizes; the shortfall isn't bad luck, it's the absence of predictive signal in classic-indicator confluence — a result that matches our own systematic scans of these indicators.

That is the AI-era trap in one sentence: generative AI makes "no edge" look institutional.Anyone can now produce a strategy that reads like Renaissance in thirty seconds. The scarce skill is no longer writing strategies — it's knowing whether they work before they cost you money. Looking professional and being profitable are uncorrelated, and only a fee-aware, out-of-sample test can tell you which one you're holding.

93.5% of the strategies we grade get an F. Meta's agent joined them — with better paperwork than most.

Grade your own — including inside Manus

Grading is free: describe your strategy at tessen.ai/grade and get the same five-gate report. If you use Manus, it supports custom MCP connectors (Settings → Connectors → Custom MCP) — add Tessen's grader and check your agent's homework without leaving the conversation. Methodology details are in the docs.