Backtesting & strategy

Strategy laboratory

Where agents and humans collaborate: an explicit strategy lifecycle, immutable versioning, and clear agent boundaries.

The strategy laboratory is where agents and humans collaborate. Every strategy moves through an explicit lifecycle, and every change creates a new immutable version.

The lifecycle

Idea → HypothesisA testable claim
Formal specificationEntry/exit rules, feature requirements
Historical test → Robustness testBacktest + out-of-sample
Shadow mode → Paper tradingLive data, no/sim capital
Human approvalGovernance gate
Limited productionVery small capital
Evaluation → Promotion or rejectionMeasured, then decided

What a strategy version pins

An immutable version captures everything needed to reproduce a trade:

Pinned	Examples
Logic	Entry rules, exit rules, scoring weights
Inputs	Feature definitions, data requirements, model versions
Risk	Risk policy, position sizing
Scope	Supported exchanges and chains
Status	Draft → testing → paper → production → rejected

The common strategy interface

Every strategy implements one interface, so the engine can run them interchangeably:

strategies/base.py

class TradingStrategy(Protocol):
    strategy_id: str
    strategy_version: str

    def evaluate(self, context) -> StrategySignal | None:
        ...  # returns a signal candidate — never places an order

It must declare supported markets and timeframes, required history and data sources, parameters, entry / exit / invalidation conditions, risk assumptions and known limitations.

The reference strategy — Momentum Breakout Spot V1

The first lab is not meant to guarantee profit; it exists to validate the whole chain end-to-end. Scope: Spot, long-only, liquid quotes, no leverage, no shorting.

momentum_breakout_spot/v1

Entry — a conjunction of measured conditions

C_{t} > H H_{20} \land \frac{V _{t}}{V _{20}} \geq 2 \land R O C_{10} > 0 \land \frac{R O C _{10}}{σ _{20}} \geq τ

Close breaks the prior 20-candle high, volume confirms (≥2× median), momentum is positive and normalized momentum clears the threshold τ — plus spread, slippage, liquidity and data-quality gates.

atr_stop_v1

Stop & size

S t o p = E n t r y - k A T R_{14} Qt y = \frac{E q u i t y \times R i s k P c t}{E n t r y - S t o p + S l i p _{u ni t} + F e e _{u ni t}}

ATR stop (optionally tightened to the breakout level), then risk-first sizing bounded by every capital, liquidity and exchange limit.

Exit on any of:

stop-loss
take-profit
trailing stop
close below breakout
momentum fades
volume collapses
spread/liquidity emergency
max holding time
circuit breaker

Agent boundaries

	Agents
May	Propose strategies, create hypotheses, suggest features, generate experiments, interpret results, recommend promotion/rejection.
May not	Silently modify a production strategy, change capital limits, activate execution, delete failed experiments, overwrite historical results.

Immutability is the point

Strategies and configurations used by a trade stay immutable, so every trade remains reproducible from its inputs and every score references the exact feature and strategy versions that produced it.