manifoldbt Backtester Skill

Purpose

Execute what backtest-expert teaches. That skill grades a backtest on five dimensions, and its prerequisites say "metrics are user-provided": it scores numbers it never produces. This skill produces them. It runs a strategy over real bars and returns the eight inputs its evaluator asks for.

The two chain in one direction: spec, run, evaluate.

When to Use This Skill

• A user describes a rule and wants it measured
• backtest-expert is about to run and the numbers do not exist yet
• A win rate, average winner, average loser or drawdown must come from bars
• A strategy's parameter count must be established for scoring

Leave the verdict to backtest-expert. It owns the thresholds and the red flags, and this skill does not duplicate them.

Prerequisites

• Python 3.9+
• pip install manifoldbt (Apache 2.0 with Commons Clause; the free tier covers

everything this skill does)

• OHLCV bars as CSV or Parquet with columns timestamp, open, high, low, close, volume
• No API key required

Workflow

1. Write the strategy spec

A spec names indicators and one entry condition. Keep it to the smallest rule that states the hypothesis. Every added knob makes an in-sample fit easier to reach by accident, and the evaluator penalises the count.

{
  "name": "sma_cross_costed",
  "indicators": {
    "fast": { "type": "sma", "period": 20 },
    "slow": { "type": "sma", "period": 60 }
  },
  "entry": { "left": "fast", "op": ">", "right": "slow" },
  "size": 1.0,
  "stop_loss_pct": 1.5,
  "fees_bps": 5.0,
  "slippage_bps": 2.0
}

Field reference: references/strategy_spec.md.

Set feesbps and slippagebps to realistic values before you read any result. A frictionless run scores 0 on execution realism, and over short holding periods costs decide whether an edge survives.

2. Run it

python3 scripts/run_backtest.py \
  --spec strategy.json \
  --data bars.csv \
  --symbol BTCUSDT \
  --json-out result.json

The script validates the spec before it touches the data, so you see a spec mistake in a second instead of after a long load.

3. Read the warnings before the numbers

The run prints warnings that change how you should read the result: a sample under 30 trades, a span under a year, no friction modelled, or a gap between the engine's win rate and the paired one. Each one is a reason to fix the setup and run again.

Three conditions stop the handoff instead of producing a score: no completed round trips, missing or non-finite maximum drawdown, and scratch trades. The evaluator has no scratch input, so passing a population that contains them would make its derived expectancy disagree with the completed trades.

4. Hand off to backtest-expert

The run ends with a command you can paste. Run it, or invoke the backtest-expert skill with the same figures:

python3 skills/backtest-expert/scripts/evaluate_backtest.py \
  --total-trades 3854 --win-rate 20.24 \
  --avg-win-pct 0.2917 --avg-loss-pct 0.2342 \
  --max-drawdown-pct 99.2893 --years-tested 0 \
  --num-parameters 3 --slippage-tested

Four conversions that fail without an error

Between an engine's output and the evaluator's inputs sit four conversions. Each one yields a plausible number and scores the strategy wrongly. None of them raises.

A fill is one execution, a round trip is two. The raw trade count runs at about twice the number of round trips. Feed fills to the sample-size dimension and you double the apparent sample, which can lift a thin backtest over a threshold it should not clear.

Buy and sell alternate only in the simplest case. That holds for a single-symbol long-only strategy that never scales a position. Shorting breaks it, because a sell can open. Scaling breaks it, because one exit answers several entries. A universe breaks it, because fills interleave. This skill tracks position per symbol and closes a trip when it crosses back through flat. Entry and exit quantities and cash values accumulate across that whole lifecycle; their weighted-average prices are display values, while PnL comes from the cash flows themselves.

Costs decide small trades. At 7 bps a side, a trade that gains 0.1% on price loses money. Expectancy comes from the win rate and the average winner together, so a gross win rate beside net averages misstates the edge. Percentages here are net of fees, and grossreturnpct sits alongside for inspection.

The engine signs drawdown negative. The evaluator wants a positive magnitude. Pass the raw value and a 38% fall scores as a flawless run.

Scope

Supported: sma, ema, rsi over any OHLC column; one entry condition using >, <, >=, <= against another indicator, a price column or a number; optional stop-loss and take-profit; fees and slippage in basis points; long-only.

Refused: multi-condition entries, shorting, multi-asset universes, and indicators outside the three above. The engine does all of these. This skill covers the shapes a one-sentence hypothesis produces, and rejects the rest instead of half-handling it.

Reference Files

• references/strategy_spec.md covers every spec field, its default, and what

validation refuses

• references/metric_bridge.md covers the eight inputs, how each is derived,

and the trap in each conversion

Scripts

• scripts/run_backtest.py runs a spec against bars
• scripts/spec.py validates a spec and counts its parameters
• scripts/round_trips.py pairs fills into round trips with net returns
• scripts/bridge.py assembles the evaluator's eight inputs

spec.py, round_trips.py and bridge.py carry no dependencies and import without the engine, so you can test the logic without running a backtest.