Advanced course

Define 1R as the planned price-risk amount before entry, then record realised net R after spread, commission, financing and slippage. Arithmetic expectancy is the mean net R per eligible trade: win rate × average win plus loss rate × average loss.

Always report count, average and median win or loss, dispersion, best and worst trade, drawdown and time exposed. Two samples can share a win rate while having opposite expectancy.

A confidence interval describes the long-run coverage of a method, not the probability that this one computed interval contains the unknown mean. State the method, resample count, random seed and interval level.

An ordinary trade-by-trade bootstrap assumes observations are sufficiently exchangeable. If results cluster by week, event or regime, test block or cluster resampling too; resampling cannot create market states absent from the data.

Skew, fat tails and clustered losses make an average fragile. Report contribution from the best one, five and ten trades, pair/year leave-one-out results and start-date sensitivity.

Reordering the same outcomes leaves total R unchanged but can transform drawdown and losing streaks. Path analysis is a risk exercise, not new performance evidence.

Write the universe, timestamps, signal, execution, risk, costs, exclusions and reject criteria before testing. Preserve raw data and a trial ledger.

Development data may shape a rule. Validation challenges it. A confirmation set is genuinely untouched only once; repeated peeking or tweaking turns it into development data and requires a new confirmation period.

Run a predeclared grid around the proposed setting and show every cell under identical costs. A narrow optimum suggests the rule may fit historical noise; a broad region of similar outcomes is less fragile but still not proof.

Repeat sensitivity across pairs, start dates, feeds and cost assumptions. Report the weakest reasonable variant, not only the winner.

A search across four entries, five exits and three filters already creates 60 variants. Informal chart checks, exclusions and abandoned ideas are also trials even if no file was saved.

Under simplified independent null tests at a 5% threshold, 100 tests produce about five false positives on average. Use a trial ledger, tougher evidence thresholds, simple baselines and methods such as deflated Sharpe or PBO where appropriate; none rescues poor data.

Long EUR/USD means long EUR and short USD; short USD/JPY means short USD and long JPY. Map every open and planned trade before adding risk percentages.

Notional exposures in different currencies require conversion to one reference currency. Stop-risk percentages are loss budgets, not notionals or volatilities, and cannot be substituted for them.

Estimate return correlation with a stated window, frequency and sign convention, then show uncertainty and alternative windows. Direction-adjust returns when asking whether current positions lose together.

Correlation does not capture every tail dependency and can change with regime. Pairwise matrices also miss a shared driver such as broad USD repricing, so add scenario tests.

Portfolio heat is entered loss if all stops fill as assumed. Add realised losses, open risk and planned risk; do not let realised gains automatically enlarge the loss budget unless the written plan says so.

Set caps for total heat, each declared factor group and event overlap. Add a stressed-cost or gap allowance and define the point at which no new trades may be added.

Record whether each trigger and fill uses bid, ask, mid or last. Model spread by session and event, round-turn commission, slippage, financing and account-currency conversion.

Report a base case and pessimistic cost case. Costs are distributions: one global average can hide the exact news or rollover period in which the strategy trades.

If an OHLC bar touches entry, stop and target, bar data may not reveal the order. Use finer data, a predeclared conservative assumption or mark the observation unresolved; never pick the profitable sequence retrospectively.

Signals based on a closing value cannot fill at that same close unless the information and execution timing make that possible. Next-bar conventions reduce look-ahead but introduce delay and gap risk that must be measured.

Store source, timezone, daylight-saving treatment, symbol specification, missing-bar checks, outlier policy, code version, dependencies and random seeds. Keep raw data immutable and derived data rebuildable.

Shadow-test on demo, compare requested and filled prices, log rejections and incidents, and define kill and rollback criteria. If a bug is found, quarantine affected results, version the fix and rerun; do not patch a report manually.

A regime rule must name the inputs, lookback, threshold, update time and execution delay. Compute each state using only data available before the decision.

Keep the state separate from the entry signal. Otherwise a filter can accidentally encode the outcome it is meant to explain.

Report always-on and filtered results with the same data, execution and costs. Include trades removed, time active, missed winners, avoided losses and the uncertainty created by a smaller sample.

A higher conditional mean can come from chance or threshold search. Preserve a locked confirmation sample and record every threshold tried.

Frequent state changes create turnover and discretionary overrides. If using hysteresis, cooldowns or minimum dwell times, predeclare and test them as parameters.

Monitor state frequency and conditional results, but do not retune at every drawdown. Define review dates, warning bands and a kill condition before deployment to demo.

Shuffling without replacement preserves every outcome and final total but changes sequence. Bootstrapping with replacement changes which outcomes appear and assumes the sample represents future draws. Block bootstrap preserves some clustering.

Parametric or synthetic paths add distributional assumptions. None creates unseen structural breaks, broker failures or gap sizes unless explicitly modelled.

For every path record maximum peak-to-trough drawdown, longest loss streak and recovery duration. Publish median and tail quantiles plus the worst simulated path and random seed.

A percentile describes paths generated by this model and sample. It is not a confidence guarantee and can be badly optimistic when the source sample misses a regime.

With fixed fractional assumptions simplified away, multiply drawdown in R by planned percent per R for a first scenario, then add compounding, gap, spread and model-error allowances separately.

Do not reverse-engineer risk to sit just inside a broker or prop-firm limit. Use a lower internal limit and a kill rule because simulated quantiles are conditional on incomplete history.

Record the contracting entity, jurisdiction, simulated or live treatment, fee and refund terms, KYC, platform, payout conditions and termination clauses for the exact checkout configuration. “Funded” does not by itself mean the trader controls live capital.

Regulatory status and consumer protections depend on entity, activity and jurisdiction. Check official registers and warning lists; a polished site or affiliate relationship is not verification.

For each rule, calculate its equity floor using the stated reference: start-of-day balance or equity, initial balance or high-water mark. Current headroom is equity minus the highest applicable floor.

Confirm reset timezone, whether floating P/L, commissions and swaps count, when a trailing floor updates, and whether it stops trailing. Never import assumptions from another programme.

Map news, overnight or weekend, EA, copier, consistency, minimum-day and maximum-open-risk rules against the strategy's evidence. A mismatch is a reject condition, not a prompt to improvise a new strategy during evaluation.

Add planned stop risk, realised loss, floating loss and a gap or cost buffer. Payout targets can encourage risk escalation; internal limits should remain independent of a deadline or fee already paid.

Include question, rationale, universe, raw-data provenance, time handling, frozen rules, execution, costs, risk, exclusions, trial ledger, code and dependencies, seeds, development and confirmation split, and the full result distribution.

Hash or version inputs and outputs. A reader with authorised access to the same data should be able to rerun the pipeline and obtain the same figures; a different sample is a separate replication test.

Use a simple baseline, negative control, cost stress, parameter neighbours, start-date and feed checks, best-trade removal, pair/year leave-one-out and locked confirmation. Record failure before interpretation.

An adverse result is useful evidence. Do not rewrite the hypothesis and keep the version number; revision creates a new candidate and requires new unseen evidence.

Allowed decisions are reject, revise as a new version, observe further, or shadow and demo test. Define data-health checks, execution drift, risk-limit breaches, performance warning bands, review schedule, kill switch and rollback before demo deployment.

Demo fills can expose operational errors but cannot validate future profitability. Any material rule change returns the candidate to research rather than being patched into a running process.