QTSurfer beta
Documentation

Metrics reference

Every field a backtest, sweep, or walk-forward result reports — its definition, units, how the engine computes it, and when it is present.

Results carry two families of numbers with different conventions. Fields whose name ends in Percent or Pct are percentages on a 0–100 scale. Fields that are rates by name — winRate, cagr — and the risk-adjusted ratios are plain ratios: 0.15 means fifteen percent, 0.58 means fifty-eight percent. Nothing is scaled twice, and nothing carries a percent sign.

All metrics are net of fees as configured on the run, and all are computed from closed trades: a position still open when the session ends contributes nothing.

Single-run results

Present in results once the strategy has closed at least one trade.

FieldUnitsDefinition
pnlTotalQuote currencySum of realised profit and loss over every closed trade.
pnlTotalPercentPercent, 0–100 scale (negative when losing)pnlTotal as a percentage of the initial capital. 0 when the initial capital is 0.
totalTradesCountClosed trades. This is the sample size behind every other metric.
winRateRatio, 0.0–1.0Share of closed trades whose profit is strictly positive. A break-even trade counts as a loss.
sharpeRatioRatioMean of the per-trade percentage returns divided by their standard deviation. See below.
sortinoRatioRatioMean of the per-trade percentage returns divided by their downside deviation. See below.
cagrRatioCompound annual growth rate of the equity, from the first to the last trade. See below.
maxDrawdownQuote currencyLargest peak-to-trough decline of the equity curve.
maxDrawdownPercentPercent, 0–100 scaleThe same decline relative to the peak it fell from.
iopsOperations per secondThroughput of the run: instrument operations processed per second. A performance figure, not a strategy metric.

How the ratios are computed

Per-trade returns. Each closed trade contributes one return, its realised profit as a percentage of the capital committed to it, at full precision. The Sharpe and Sortino ratios are statistics of that series, not of calendar-period returns:

sharpeRatio  = mean(r) / stdev(r)
sortinoRatio = mean(r) / downsideDeviation(r)
  • Neither ratio is annualised, and neither subtracts a risk-free rate. Compare them with other QTSurfer results, not with annualised daily figures from elsewhere.
  • Both report 0 when fewer than two trades were closed. sharpeRatio reports 0 when every trade had the same return; sortinoRatio reports 0 when no trade lost money — a strategy without a losing trade shows a zero Sortino, not an infinite one.
  • stdev is the population standard deviation of the series.

CAGR. Computed from the equity path, not from trade returns:

years = (lastTradeTime − firstTradeTime) / 365.25 days
cagr  = (finalEquity / initialCapital) ^ (1 / years) − 1

Windows shorter than one year are deliberately not annualised: extrapolating a few weeks to a year produces meaningless magnitudes, so for years < 1 the field equals the plain equity return, finalEquity / initialCapital − 1, which is the same quantity pnlTotalPercent reports on its own scale. The two formulas agree exactly at one year. cagr is 0 when there is no positive initial capital, no elapsed time between trades, or the equity went to zero or below.

Drawdown. Tracked on the equity curve as trades close: the running peak, and the deepest fall from it in currency and as a fraction of that peak. See the drawdown article for the calculation and its reading.

Equity curve

results.equityCurve is present under the same condition as the metrics. Its first point is an anchor at the backtest start with the initial capital; each later point is recorded when a trade closes, so equity = initialCapital + cumulativePnl at that moment. Points are account value in quote currency, never percentages. Shapes, transforms, and the metadata that says what was actually served are documented in Equity curves.

Notices

results.notices lists diagnostics the engine raised, each with a level, a code, a message, and a provenance. The field is absent when nothing was raised — it is the one place where silence is a real answer. Notices are also raised on failed and aborted runs, and a run with no trades usually explains why here. noticesTruncated reports how many were dropped past the cap of fifty, and is absent when none were.

Sweep leaderboard rows

Each trial in a sweep reports its own copy of the run metrics plus the fields that make trials comparable:

FieldUnitsDefinition
runIxIndexDeterministic position in the expanded grid; stable across shards and rankings.
rankPositionPresent only in the ranked view.
paramsMapThe parameter vector of this trial.
sharpe, sortino, pnl, pnlPct, cagr, maxDdPct, trades, winRateAs aboveThe trial’s own results. pnlPct and maxDdPct are on the 0–100 scale; cagr and winRate are ratios.
plateauScoreSame units as the objectiveThe objective of the worst run in this point’s immediate neighbourhood. The default ranking key.
neighbourCountCountNeighbours that existed for the plateau score. 0 means the score is unevidenced, not confirmed. Read the two together.
deflatedSharpeProbability, 0.0–1.0Probability that this trial’s Sharpe reflects real edge rather than the best draw among the vectors tried, accounting for sample length, skewness, and kurtosis. Above roughly 0.95 survives the correction; at or below 0.5 is indistinguishable from the best of a pile of random trials. Absent, not zero, when the statistic isn’t meaningfully computable — an aborted run, too few trials to establish any dispersion, fewer than three period returns, or a near-constant return series.
belowTradeFloorBooleantrades fell below the sweep’s minTradeFloor (thirty by default). The row stays in the results.
abortedBooleanThe run threw and measured nothing. Aborted rows are excluded from sensitivity aggregates.
runtimeMsMillisecondsWall-clock time of the trial.

At sweep level, pbo is the probability of backtest overfitting over the whole grid, from combinatorially symmetric cross-validation, with pboSplits the number of splits used. Above roughly 0.5 the selection is picking noise. Both are present only once the last shard finishes, and only for a sweep without walk-forward.

The sensitivity endpoint aggregates the objective per parameter value as best, mean, and worst, with count runs behind each point. It excludes aborted runs, and it can be read while the sweep is still running.

Walk-forward results

A walk-forward sweep reports one row per completed fold instead of one per parameter point:

FieldDefinition
foldIxPosition of the fold, oldest first.
inSampleFrom, inSampleTo, outOfSampleToWindow indices into the prepared session.
paramsThe vector that won the fold’s in-sample optimisation.
inSampleSharpeThat winner’s Sharpe on the window it was chosen on — the flattering number, there to compare against the next field.
outOfSampleA full leaderboard row scored on the following, unseen window — the honest number.
vectorsRunVectors evaluated in-sample before picking the winner.

paramDrift, at the sweep level, is the mean normalised lattice distance between consecutive fold winners. It is omitted, not zero, until at least two folds have completed, because zero is itself a meaningful reading. Walk-forward sweeps report no plateau score, deflated Sharpe, or PBO: those corrections exist to deflate in-sample winners, and out-of-sample scores need no deflating.

Reading the numbers together

  • totalTrades first. Every ratio is a small-sample statistic when the trade count is small, and belowTradeFloor exists to make that visible in sweeps.
  • sharpeRatio with winRate, maxDrawdownPercent, and the equity curve. A high ratio on a curve that made all its money in one episode is a story about that episode.
  • cagr with the session length. Under one year it is a plain return, not a growth rate.
  • deflatedSharpe and pbo before rank. The top of a leaderboard is where selection bias lives.