How to Evaluate Live Algorithm Performance

Create a provenance record from the published methodology. Determine whether the page reports a model, simulated account, vendor account, customer accounts, or aggregated signals. Note who controls the account and whether records can be reconciled to order or broker data. If those facts are unavailable, downgrade confidence rather than filling gaps with assumptions.

Mixed histories need clear segmentation. A strategy can have a long hypothetical backtest and a short live period, but the two should not be joined without visible labels. Use the live segment to study observed execution and behavior; use the backtest to examine broader historical hypotheses. Neither segment guarantees future performance.

Futures leverage complicates denominator choices. Account balance, required margin, notional exposure, and intended risk are different quantities. Margin is a performance bond and can change; it is not maximum loss. State which capital measure a percentage uses, and do not compare percentages built from different denominators.

Costs must follow each trade. Include commission, exchange and regulatory fees where applicable, and actual or conservatively modeled slippage. If a public result is gross, do not silently compare it with a net alternative. Estimate a range and retain the calculation. A high-turnover strategy can be especially sensitive to small cost changes.

Win rate needs average win, average loss, and exit behavior. A high win rate can coexist with rare severe losses; a lower win rate can accompany positive asymmetry. Review maximum favorable and adverse movement when available, but remember those are sample statistics rather than future bounds.

Plot outcomes by time, market, direction, session, and strategy version. Look for serial dependence and bursts of activity that reduce the effective sample. If many trades are near-identical or occur in one regime, the nominal count overstates independent evidence. Confidence should reflect diversity of observations, not a round number.

Plot the underwater curve rather than reading only the maximum value. Two strategies with the same maximum drawdown can differ materially if one recovered quickly while the other remained below its peak for months. Long stagnation affects capital opportunity, operator behavior, and whether a strategy remains within external account rules.

Identify the mechanism of severe losses. Were they concentrated in a volatility shock, repeated mean-reversion entries, an overnight gap, a contract roll, or execution failure? A causal hypothesis helps design a relevant stress. It does not justify deleting the period or assuming the next event will look identical.

Ask when the strategy should struggle. A credible mandate includes unfavorable conditions and an explanation of how risk is controlled. If no losing regime can be identified, the description may be too vague to test. Compare that expectation with actual drawdowns and avoid rewriting the mandate after observing results.

Selection is a major source of overconfidence. A catalog may show surviving or currently promoted systems while failed candidates disappear. Ask how additions, retirements, and revisions are recorded. Evaluate the selection process and preserve the full candidate set where possible, not just the best result chosen with hindsight.

Do not convert historical metrics directly into expected income. The CFTC cautions against claims that trading bots or systems can guarantee returns, and NFA rules explain limitations of hypothetical performance. Treat every historical record as uncertain evidence. Size from loss tolerance and operational capacity, not from a target payout.

Set review triggers for strategy revisions, unexplained divergence, cost drift, drawdown, regime change, and data-quality problems. Compare live behavior with the frozen thesis, but allow for statistical noise. Changes should be versioned and re-evaluated; quietly replacing settings destroys the continuity that made the record useful.