Intermediate course

A pivot rule has a confirmation delay. If two bars on the right are required, the swing cannot be known on the swing bar itself. Mark both the pivot time and the time it became observable. This prevents a backtest from acting on a level before the rule could have identified it and exposes how much apparent precision comes from hindsight.

Define whether a break requires a trade, wick, close or close plus distance beyond a level. Then define the evaluation horizon for acceptance. A quick return inside can be labelled failure only if that condition existed beforehand. Without those definitions, the same sequence can be called breakout, sweep or false break depending on the preferred story.

Top-down analysis should transfer only information that was available at the decision timestamp. Freeze the higher-timeframe state first, then inspect the lower timeframe. Record disagreements rather than forcing alignment. A lower-timeframe setup may be rejected, accepted with reduced expectations or studied as a separate category, but it should not rewrite the higher-timeframe close.

Every indicator has an input, transformation and output. A longer moving average smooths more variation but reacts later. That is not a defect to optimise away; it is the trade-off created by the formula. Inspect how a parameter changes signal timing and turnover across a broad range rather than selecting the one value that best fits the sample.

Choose an indicator by job. Trend filters classify direction, oscillators describe relative position or momentum, and volatility measures help scale stops or expectations. “Oversold” is not automatically a buy instruction, and ATR does not predict direction. Write the decision the indicator is allowed to influence and the decisions it cannot make.

Different names do not guarantee independent evidence. RSI, stochastic and MACD all transform related price history, so simultaneous agreement can be one data source counted three times. Map each feature back to its raw inputs and test incremental value: compare the full rule with an otherwise identical version that removes one indicator.

Begin with measurable geometry: pivot rule, number of touches, permitted slope, tolerance around a level and maximum duration. Name the formation only after those facts are recorded. This reverses the usual pattern-matching habit, where a memorable label encourages the analyst to ignore points that do not fit.

A candidate pattern is not a confirmed event. Define confirmation as a completed, timestamped condition such as a close beyond a boundary plus a volatility buffer. Also define entry timing: confirmation close, next open or retest. Each choice changes available price, stop distance and selection, so it must be tested as a separate rule rather than blended after the outcome.

Context and counterexamples determine whether the geometry adds value. Tag trend, volatility, session and nearby higher-timeframe levels using information known at the time. Keep failed and ambiguous candidates. If only attractive textbook examples are retained, the sample estimates how well hindsight can curate a chart, not how the rule behaves in use.

The research contract states the market, timeframe, session, data source, costs, eligibility filter, entry, invalidation, size, exit and review metrics before testing. It also states exclusions. This document creates friction deliberately: if a rule cannot be written clearly, it cannot be reproduced consistently enough for its performance numbers to mean much.

Make actions code-like even when testing manually. Replace “enter on momentum” with observable inputs, thresholds, timestamps and order assumptions. Add a no-trade output for incomplete information. Precision does not mean adding decimals; it means that two careful reviewers should reach the same decision on the same frozen chart.

Test interpretation before profit. Give a sample of charts to another reviewer, or repeat the classification after a delay without answers visible. Measure agreement on eligibility, trigger and invalidation. Low agreement means performance combines several hidden strategies. Fix the definitions before expanding the sample or tuning parameters.

Design the sample before inspecting results. Include different volatility states, sessions and market conditions relevant to the intended use. Split chronologically so later evaluation data mimics a future deployment decision. Randomly mixing neighbouring FX observations can leak regime information and exaggerate how independent the test really is.

Look-ahead bias can enter through indicators, revised economic data, session labels, swing confirmation and manual chart scrolling. For each feature, record when its value became available. An entry can use only the version known at that timestamp. Freeze ambiguous fill rules as well, because OHLC bars do not reveal the order in which high and low occurred.

The result is a distribution, not one win rate. Inspect average R, median, dispersion, losing streaks, drawdown, turnover, time in market and concentration by year or regime. Add costs and stress them. A positive average driven by a few extreme trades requires different operational confidence from a broad, stable distribution.

A central-bank reaction function is a framework for how officials may balance inflation, activity, employment and financial conditions. It is not a formula that maps one data release to one trade. Track the language, forecasts and distribution of committee views, then state what policy path the market appeared to price before the decision.

Markets react to surprise relative to expectation and to changes in the expected future path. A rate rise can accompany currency weakness if it was fully priced or paired with softer guidance. Separate the announced fact, prior consensus, immediate repricing and later interpretation. This prevents a post-event explanation from being treated as a rule available beforehand.

FX is relative. A constructive domestic story may not strengthen a currency if the other side of the pair improves more or carries a stronger expected-rate path. Build a two-column comparison covering growth, inflation, policy, external balance and risk sensitivity. Then define which observable change would revise the comparison instead of defending a permanent bias.

Replace identity labels such as impatient or fearful with observable behaviour: moved a stop after entry, entered before confirmation, exceeded the daily loss budget or skipped the review. Add time, context and trigger. Behavioural data becomes useful when another reviewer can recognise the same event without interpreting your personality.

An if-then safeguard links a trigger to a pre-committed action. If two planned losses occur, then pause new entries until the next session review; if an unscheduled high-impact event is near, then cancel the pending order. The safeguard must be practical under stress and should reduce available choices rather than ask for more willpower.

Review adherence before returns. Score whether the setup was eligible, the plan was frozen, risk was within budget and execution matched the rule. Then inspect P/L. This order prevents a lucky deviation from being rewarded and a well-executed loss from being rewritten as a psychological failure. Track repeated deviations to decide whether the process or environment needs redesign.

A playbook is more than a screenshot collection. For each setup, include eligibility, trigger, invalidation, sizing, exit, costs, examples, counterexamples, sample summary, operating checklist and source notes. A reader should be able to trace a live decision back to the exact tested rule and the evidence supporting it.

Acceptance and rejection criteria belong in the playbook before forward testing. Define tolerances for execution cost, rule adherence, opportunity frequency and deterioration, plus hard safety limits. Avoid making one short losing sequence an automatic rejection or one winning sequence an approval; use the uncertainty and drawdown observed in research to set proportionate review gates.

A forward test runs the frozen version in sequence without restarting whenever results disappoint. Log missed and rejected trades as well as entries. If a material rule changes, assign a new version and preserve the old evidence. This keeps learning cumulative: otherwise every adjustment erases the denominator and produces an endless first sample.