Captain · Independent Research and Operations · About 30–40 minutes
Safety rule: A research hypothesis is not a trading edge until it survives independent cost-aware testing.
Your mission today
By the end you can turn a vague market belief into a falsifiable question with a defined sample and rejection rule.
Before you start
Complete the Captain strategy and research modules. Bring your frozen strategy card and a small synthetic EURUSD dataset; avoid using live capital.
1. Replace a clever story with a testable claim
A weak market claim sounds persuasive but cannot be checked. 'EURUSD respects strong levels' contains no definition of strength, respect or success. A research question identifies a measurable event, an outcome and conditions under which the idea could be rejected. Think of testing a new medication: you must define who qualifies, what is measured and what counts as failure. In trading, that means defining symbol, timeframe, decision timestamp, qualifying chart structure, entry, protective stop, exit and transaction costs. Use completed candles only when your rule requires closed bars. State the account and contract assumptions separately from the pattern. A hypothesis is a proposed explanation or relationship that may be false. Strong research actively seeks conditions that would disprove it. Do not begin by scrolling to the most dramatic winning candle and designing a rule around that one success. Decide the sample dates before reviewing outcomes.
2. Write the downside into the hypothesis
Imagine a synthetic rule: after an H1 candle closes above a premarked EURUSD range high, a hypothetical long may be eligible at the next Ask if spread and risk limits allow. The invented entry is Ask 1.08520 with a planned Bid stop 1.08420. That is 10 pips of planned downside. At a fictional $1 per pip, the price loss would be −$10 before costs and worse fills. A later executable Bid at 1.08670 would be +15 pips, or +$15 gross, in one possible outcome. But a different path may stop first. The research question is not 'Can I earn 15 pips?' It is whether the defined event reaches a specified executable target before its stop, frequently enough to cover realistic costs. Unknown fills and ambiguous stop-versus-target paths need predeclared handling rules.
3. Define the population and the clock
The population is the complete set of events your question intends to describe. For example, every eligible H1 breakout within a fixed three-month study interval is a population for that interval. A sample is the portion you actually observe or analyse. If you select only breakouts that later travel far, the sample is biased. Start with an event counter and record every offered candidate, including those rejected because the spread exceeded the rule. Identify which clock defines a close. Broker server time may not equal Kuala Lumpur time, and session boundaries can shift with daylight-saving rules. A qualifying pivot requiring two right-hand bars is not knowable at the pivot's own timestamp. Your research ledger should contain both the pivot bar and the first time it was confirmed. Otherwise an impossible early entry may make the hypothesis appear more successful than it is.
4. Separate development from independent evaluation
Use a development sample to design and choose candidate rules. Reserve a later holdout sample that is not inspected while choosing them. If you repeatedly tune a threshold on holdout results, that sample has become development work too. A rejection rule states in advance what result would make you abandon or revise the hypothesis. It might require a certain net expectancy threshold after modelled spreads, but the threshold itself should be grounded in the intended application and measurement uncertainty. Do not choose it after seeing the answers. Include all trades, skips, losses and costs. Keep code and data versions so another reviewer can reproduce the outcome. Be cautious when multiple currency pairs share the same event. Treating ten correlated trades as ten independent confirmations can exaggerate confidence in a small underlying sample.
5. Example: compare a claim with its evidence
Suppose a fictional dataset contains 40 chronological candidate breaks and 18 reach the declared target before the stop. The observed fraction is 18 ÷ 40 = 45%. That number is an event rate for the toy sample, not a forecast of future trades. If the target offers +1R and the stop loses −1R before costs, 18 wins contribute +18R and 22 losses contribute −22R, totalling −4R, or −0.10R per offered completed trade under this simplified binary outcome model. Added costs make the result worse. Notice that a chart gallery of 18 successful breaks could look impressive while the full forty-event ledger is unprofitable. Document any rejected or unfilled opportunities separately rather than quietly changing the denominator. Do not conclude that a larger profit target automatically fixes the low hit rate. The actual test needs new evidence and consistent exit assumptions.
6. Produce a pre-registration sheet
A pre-registration sheet records the hypothesis before outcomes are examined. Write the precise instrument, contract, timeframe, sample start/end, observable signal, first knowable time, entry and exit quote sides, costs, ambiguous-fill policy, cash-risk limit, exclusion rules and pass or rejection criterion. Next record the development/holdout split and how many rules you intend to compare. An audit trail is the saved set of decisions and inputs that lets someone else reproduce the test. Give your sheet to a second reviewer and ask them to identify one event that would qualify, one that would be rejected and one that would remain unknown. If they interpret your conditions differently, refine the rule wording before starting. The final research memo must distinguish arithmetic, observed sample evidence and speculation. A credible Captain should be willing to report 'no dependable edge found' when the data does not support a positive claim.
Common mistake
A researcher starts with winning EURUSD examples and then writes a rule that explains all of them. The losing candidates never enter the denominator. Fix the method by defining eligibility and sample boundaries before looking at results.
Plain-English glossary
- Research question: an observable claim with measurable acceptance and rejection.
- Hypothesis: a proposed relationship that could be false.
- Population: all events covered by the defined study.
- Holdout: data reserved from rule selection for later evaluation.
- Audit trail: saved records allowing independent reproduction.
Try it yourself
Write a one-page pre-registered EURUSD research plan. Include ten mandatory fields, the complete event population, a loss-first example, cost assumptions and an independent holdout. Ask a reviewer to classify three hidden-future cases using only your plan.
What you can now do
- Explain which evidence supports the written claim.
- Reproduce the loss or quality calculation with correct units.
- Deliver a dated, independently reviewable research record.
Next lesson
CP07.2 · Data Quality: Can You Trust Your Chart Data? Check whether the evidence itself deserves confidence.
