Three checkpoints in this guide
Follow the full walkthrough in order, or jump directly to one of its main sections.
A profitable month does not prove a durable strategy, and a losing month does not prove that a valid process failed. P&L is an outcome assembled from position size, opportunity mix, market path, execution, costs, and luck. Performance analysis starts by reconciling those inputs, preserving their time order, and asking how uncertain each estimate is.
The ten metrics below do not “predict success” one by one. They form a diagnostic system. Expectancy without uncertainty can be noise; win rate without payoff is incomplete; drawdown without exposure is hard to interpret; a strong backtest selected from many trials may be overfit. Use the metrics together, against a declared strategy version and a forward decision rule.
The short answer: analyze net outcomes in a consistent unit, then add path, risk, execution, cost, and stability measures. Show the estimate, the sample and time window, uncertainty, missing data, and every filter used. A metric becomes actionable only when it maps to a decision that was specified before the next observation.
What Trading Metrics Can—and Cannot—Tell You
A metric is a compressed description of recorded trades. It cannot identify a cause by itself. Lower performance in one session may reflect different setups, volatility, instruments, news exposure, or a small number of outliers. “Stop trading that session” is therefore a hypothesis to test, not an automatic conclusion.
Before calculating anything, lock five definitions:
- Strategy version: exact entry, exit, sizing, eligibility, and management rules.
- Observation unit: order, fill, round trip, campaign, trade idea, or daily return.
- Outcome unit: account currency, return, basis points, or R based on declared initial risk.
- Clock: decision, order, first fill, final fill, or market-session timestamp and timezone.
- Cost policy: commission, spread, slippage, financing, borrow, funding, conversion, data, and other applicable charges.
Reconcile source totals before grouping trades. Preserve cancelled orders, partial fills, deposits, withdrawals, corrections, and missing fields. A clean-looking dashboard built on incomplete records is less useful than a visibly incomplete report marked Not verified.
10 Trading Metrics to Read as One System
1. Net Expectancy
Net expectancy is the sample mean of outcomes after applicable costs. For binary win/loss summaries it can also be written as win probability times average win minus loss probability times average loss, but the direct mean retains scratches and other outcomes without forcing them into two buckets.
Net expectancy = Σ net outcomeᵢ ÷ N.
Report expectancy in both a normalized unit and account currency when possible. R is useful only if initial risk was defined consistently and changes to stops or size remain recorded. Currency is useful for reconciliation but can mix different exposure. The expectancy formula guide shows how the pieces connect.
Do not multiply a noisy sample mean by a future trade count and call the result expected income. Show dispersion, time order, strategy version, and an uncertainty interval appropriate to the data. If trades overlap or cluster by market regime, treating every trade as independent can overstate precision.
2. Win Rate With Payoff Distribution
Win rate is the share of eligible observations classified as wins under a fixed rule. It is not quality by itself. Two strategies with the same win rate can have different average losses, tails, costs, exposure, and paths; two strategies with different win rates can have similar expectancy.
Win rate = winning observations ÷ eligible observations.
Always pair it with average and median win, average and median loss, scratches, largest gains and losses, and quantiles of the full distribution. Define whether a partial exit creates one trade or several fills. Do not compare a fill-level win rate with a campaign-level payoff ratio.
A rising win rate can be bad if profits are cut earlier while losses retain their full size. A falling win rate can be acceptable if the payoff distribution changed by design and net expectancy remains supported. The metric describes frequency; the strategy definition explains what frequency means.
3. Profit Factor
Profit factor compares total positive net outcomes with the absolute total of negative net outcomes for the same sample.
Profit factor = gross net profits ÷ |gross net losses|.
Values above one describe a profitable historical sample before uncertainty is considered; they are not universal evidence of a “good” system. The ratio is undefined when the denominator is zero, sensitive to a few large winners, and easy to inflate by filtering after seeing results. Report the number of wins and losses, sample period, costs, and profit factor with the largest trades removed as a sensitivity check—not as a replacement result.
When comparing instruments or sessions, use the same strategy version and eligible population. A subgroup can look strong merely because a few favorable regimes occurred there. The profit-factor benchmark guide explains why context and uncertainty matter more than generic labels.
4. Planned Risk, Realized R, and Rule Adherence
Planned reward-to-risk is not the same as realized payoff. For each eligible trade, preserve the entry decision, initial stop or invalidation, intended management, quantity, and target before the outcome is known. Then calculate realized net R against that fixed initial-risk denominator where the data support it.
Realized net R = net trade outcome ÷ declared initial risk amount.
A gap between planned and realized R does not automatically prove fear, greed, or poor discipline. It can come from partial fills, slippage, gaps, dynamic exits that were part of the plan, platform behavior, manual intervention, or bad source data. Separate rule adherence from outcome quality: a compliant loss and a non-compliant win require different diagnoses.
Track discrete rule fields rather than one subjective score. Examples include setup eligibility, size, initial stop, add/scale rule, exit rule, prohibited window, and required evidence. An adherence rate is meaningful only if every denominator event was actually recorded.
5. Drawdown Depth, Duration, and Recovery
Maximum drawdown is the largest peak-to-subsequent-trough decline in a chosen equity series. State whether the series is balance or marked equity, gross or net, and how deposits and withdrawals are handled.
Drawdown at time t = (running peak − current equity) ÷ running peak.
Depth alone loses information. Also report time or observations from peak to trough, time under water until a new high, the distribution of drawdowns, and concurrent gross exposure. A sharp loss caused by one oversized position and a long grind across many correctly sized trades can have the same maximum depth but different remediation.
There is no universal acceptable drawdown. The decision depends on capital, mandate, leverage, investor or program rules, loss tolerance, strategy distribution, and the uncertainty of the estimate. Use the drawdown tracking guide to keep balance, equity, daily and trailing definitions separate.
6. Exposure and Risk Utilization
Returns cannot be interpreted without knowing how much and what kind of risk produced them. Track gross and net exposure, leverage or margin use, initial risk, current risk to planned exits, concentration by instrument and theme, and time spent near account or program limits.
Useful denominators include return per unit of average exposure, return per unit of declared initial risk, and cost per unit of turnover. None is sufficient alone. An apparently consistent strategy may simply operate at variable exposure, while a flat P&L period may hide a substantial reduction in risk.
For prop-style trading, “consistency” must be tied to the exact program and verified rule date. Track daily and maximum-loss distance, open-position treatment, consistency or concentration mechanics, scaling state, and prohibited practices. Do not infer a pass probability from generic percentages or use one firm’s definitions for another.
7. Tail Loss and Loss Clustering
Average loss does not show the events most likely to breach a limit or force a strategy change. Report worst trade, worst day, worst rolling window, downside quantiles, gap or slippage events, and the contribution of the largest losses. Keep operational errors separate while still including their account impact.
A maximum losing streak is a historical path statistic, not a reliable forecast by itself. Simple streak formulas usually assume independent observations with a stable win probability. Trading outcomes can be serially dependent because setups overlap, markets share shocks, behavior changes after losses, and regimes persist. Inspect lagged dependence and clusters before using an independence model.
Set response rules from account survival and evidence quality, not a claim that a particular streak is “normal.” If a boundary is crossed, freeze the affected strategy version, reconcile the data, check operational causes, and compare the event with predeclared stress cases before deciding whether to resume.
8. MAE, MFE, and Time in Trade
Maximum adverse excursion (MAE) is the worst marked movement against a trade during its life; maximum favorable excursion (MFE) is the best marked movement in its favor. Both require intratrade prices, a defined mark source and timestamp resolution. Bar data can miss the sequence inside a bar, and bid/ask treatment matters.
Use MAE and MFE to test stop, target, and management hypotheses—not to move every historical stop to the point that would have improved the same sample. Compare declared rules out of sample. Preserve trades that never reached a proposed level and include execution feasibility.
Duration should be conditioned on setup and exit type. “Winners are shorter than losers” does not reveal whether the cause is a time stop, trend capture, news, early profit taking, or selection. Report decision-to-fill, fill-to-exit, and time under material exposure separately when they answer different operational questions.
9. Cost Drag and Execution Quality
Cost drag is the difference between a consistently defined gross result and the reconciled net result. Include every applicable explicit and implicit layer for the exact market and account. The all-in trading-cost guide covers commission, spread, slippage, financing, borrow, funding, data, conversion, and missed-fill trade-offs.
Cost drag = gross strategy result − reconciled net result.
Execution quality needs a declared benchmark available at the decision time. Record order type, decision timestamp, reference bid/ask, every fill and partial, cancellations, replacements, rejected orders, latency where available, and unfilled eligible opportunities. A lower commission does not prove lower all-in cost; a lower measured cost does not prove the strategy retained the same opportunity set.
10. Stability Across Time, Regime, and Segment
Stability asks whether the result survives reasonable changes in time window and relevant conditions. Display rolling or expanding estimates, but do not treat overlapping windows as independent confirmation. Predeclare the segments that matter—strategy version, instrument, session, volatility state, direction, setup, execution route, and account type—and retain small denominators.
Multiple testing is a central risk. If dozens of sessions, indicators, thresholds, and instruments are searched, the best historical slice can look exceptional by chance. Keep a trial registry, separate exploration from confirmation, reserve untouched data or future observations, and report all tested variants rather than only the winner.
An expert advisor should be evaluated the same way: current version ID, out-of-sample and forward periods, costs, drawdown path, exposure, trade dependence, recent outcome attribution, and any parameter changes. Recent weakness is a diagnostic signal, not automatic proof of failure; a favorable backtest is not automatic proof of live robustness.
What Does Good Trading Performance Look Like?
There is no responsible table of universal “good trader” percentages. A result is decision-useful when its definition is stable, the source reconciles, the sample covers the conditions in which the strategy will be used, costs and exposure are included, and uncertainty is narrow enough for the proposed action. A positive point estimate with a wide interval is evidence to keep measuring, not permission to scale.
| Dimension | Stronger evidence | Weak evidence |
|---|---|---|
| Return source | Many eligible observations across declared conditions | A few outliers or one favorable episode |
| Risk | Exposure and tail states remain inside predeclared limits | Average return shown without path or leverage |
| Execution | Decisions, orders, fills and costs reconcile | Backtest fills or screenshots without controlling records |
| Stability | Result survives untouched time and relevant segments | Best slice selected after many trials |
| Process | Rule adherence is measured separately from outcome | Profitable trades relabeled as compliant afterward |
Comparing one trader with another is usually less informative than comparing the same declared process across versions. Capital, leverage, mandate, instruments, opportunity frequency, fees, and loss limits may differ. For a prop-style account, “good” first means staying inside the exact verified rules; for a personal account, it still requires an explicit risk mandate rather than a borrowed benchmark.
Use thresholds as operational alarms only when their origin and response are documented. For example, a team may pause a strategy when reconciled drawdown breaches its precommitted boundary. That is a governance rule, not evidence that the same percentage is optimal for every strategy. Record who set it, which equity definition it uses, when it was verified, and what must happen before resumption.
How to Read Trading Metrics Together
Start with a question, not a dashboard. Define the observation population and the action that would follow. Then look for alternative explanations and specify what evidence would reject the proposed change.
| Observed pattern | Competing explanations | Next test |
|---|---|---|
| Higher win rate, lower expectancy | Smaller winners, larger tails, more costs, changed exits | Reconcile payoff distribution and rule versions |
| Positive profit factor, deep drawdown | Concentration, serial loss, leverage, one tail event | Attribute exposure and ordered loss path |
| One session underperforms | Setup mix, instrument mix, events, small sample | Predeclare a controlled forward restriction |
| Planned R exceeds realized R | Rule breach, planned dynamic exit, fills, bad fields | Audit decision records and execution separately |
| Strong backtest, weak live result | Overfit, costs, regime, implementation, data leakage | Freeze version and reconcile model-to-live deltas |
Consider a hypothetical version with positive net expectancy but most profit concentrated in a few trades, rising drawdown duration, and negative cost-adjusted results outside one instrument. The defensible conclusion is not “stop the bad instrument” or “hold every winner longer.” First verify the outliers, dependence, exposure, costs, and whether the instrument was part of the original hypothesis. Then test one predeclared change on later data.
A Reproducible Performance-Review Cycle
- Reconcile: match fills, cash movements, fees, financing and final totals to the controlling records.
- Version: freeze strategy, account, import, normalization, timezone, cost and metric definitions.
- Describe: show all ten metrics with denominators, missingness, time order and relevant segments.
- Diagnose: write one hypothesis, plausible confounders, expected change and rejection evidence.
- Validate: test on untouched or future observations with the same eligibility and cost policy.
- Decide: keep, revise, reduce, suspend or reject the exact version; record the reason and rollback rule.
Calendar cadence should follow the strategy’s information rate. A weekly review may be useful operationally but statistically empty for an infrequent strategy; hundreds of intraday fills can still represent only a few correlated market episodes. Use quick process checks frequently and make strategy changes only when the evidence window matches the decision.
Where TSB Fits in Performance Analysis
Ownership disclosure: Trader's Second Brain is our product. It is relevant here as an import, journal, prop-rule tracking, and retrospective-analysis workflow—not as a broker, execution venue, independent auditor, or guarantee of edge, consistency, program compliance, or future performance. TSB recognizes 332 structured source profiles through canonical runtime truth.
Use the broker, venue, platform, or program record as controlling evidence. Preserve the original file; map order and fill identifiers; keep account currency, timestamps, quantities, fees and financing; declare the trade-grouping and R-normalization rules; and version every later edit. The journal field guide provides the minimum decision context that a statement usually lacks.
Acceptance test: imported totals reconcile, each metric exposes its formula and denominator, unsupported fields remain Not verified, all filters and strategy versions are visible, and a claimed insight links back to the exact observations that produced it.
Methodology and Evidence Boundaries
This guide was reviewed on September 9, 2026. NIST’s autocorrelation reference supports checking whether ordered observations are non-random rather than assuming independence. Andrew Lo’s Statistics of Sharpe Ratios derives sampling behavior under different return assumptions and shows why serial correlation can change annualized inference. Halbert White’s Reality Check for Data Snooping addresses the risk of mistaking a best result selected from repeated use of the same data for genuine predictive superiority.
These sources do not validate a retail benchmark for win rate, profit factor, drawdown, sample size, session selection, or prop performance. The prior guide’s universal thresholds, fabricated trader examples, fixed session schedules, named-program numbers, causal psychology diagnoses, automated-product guarantees, and projected income calculations were withdrawn. Production remains read-only until a separately authorized content import.
Final Verdict: Metrics Are a Decision System, Not a Scorecard
Reliable analysis keeps the raw path visible. Reconcile the data, freeze definitions, normalize outcomes, show uncertainty and dependence, add risk and cost context, and challenge every attractive segment on untouched or future observations. No single percentage proves that a trader, strategy, or expert advisor is good.
The useful output is a versioned decision: what evidence changed, what action follows, what would falsify it, and how to roll back. When those fields are missing, the right status is Not verified—not a benchmark, prediction, or story.
