Algorithmic Trading Backtest Lab

Almost every published backtest is wrong in the same direction. Not fraudulently — structurally. The default way to write one quietly includes a set of advantages the strategy would never have had:

• Prices that were revised after the fact, treated as if they were available live.
• A universe of instruments that still exist today, silently excluding the ones that failed.
• Zero transaction costs, or a flat fee that ignores spread and market impact.
• A parameter chosen by testing forty variants on the same history.

Each of these makes results better. None of them are available to a real trader. A framework that makes them easy to do by accident is producing fiction with a Sharpe ratio attached.

A research harness whose defaults are pessimistic. Costs, slippage and point-in-time universes are on by default and must be explicitly disabled; walk-forward is the standard evaluation rather than an advanced option.

The reporting is deliberately unflattering: in-sample and out-of-sample results are always shown side by side, and every report includes the degradation between them. If a strategy only works in-sample, the report says so on the first line.

This is educational research infrastructure. It is not investment advice and produces no recommendation to trade anything.

Costs are not a detail#

The most common objection to including realistic costs is that they are "roughly a rounding error". For anything trading more than a few times a month, they are not.

costs.pypython
def transaction_cost(
    notional: float,
    volatility: float,
    adv_participation: float,
    commission_bps: float = 1.0,
) -> float:
    """Commission + half-spread + a square-root market impact term."""
    commission = notional * commission_bps / 10_000
    half_spread = notional * (0.5 * volatility * 0.1)
    impact = notional * 0.1 * volatility * (adv_participation ** 0.5)
    return commission + half_spread + impact

The square-root impact term is what stops the harness rewarding strategies that only work at sizes nobody could actually trade. Without it, the optimiser reliably finds a "strategy" that trades enormous size in illiquid instruments.

What walk-forward actually shows#

Running a moving-average crossover with a parameter grid, in-sample and out-of-sample:

Roughly a quarter to a third of in-sample performance survives, and one window is negative. That is the normal, honest result for a simple technical strategy, and the fact that it looks disappointing is the finding.

Why the framework is built to disappoint#

A backtesting framework has one job: to be harder to fool yourself with than a spreadsheet. Every default in this one is chosen to make the strategy's life difficult, because the alternative — a framework whose defaults flatter — is how people end up trading a curve fit.