The last three lessons built an argument: prices move on information and expectation, some information events force constrained participants to act, and the events that are recurring, dated, and disclosed can be queried across the whole market. What remains is the procedure that turns a queryable event into a decision.
That procedure has four steps. We are going to state them once, precisely, because every hands-on lesson in the next eleven modules is another repetition of the same loop on a different dataset. By the end of the course you will have run it a dozen times, and the goal is that you become unable to think about a trading idea any other way.
Hypothesis, data, measurement, attack. Four steps, in that order, every time.
To keep this concrete, we will carry one example through all four steps: a study of earnings moves that we will actually run, code and all, in the Volatility II module. Here we only walk the reasoning.
Step 1: Hypothesis
State the claim precisely, before touching data.
A hypothesis is a single sentence with a yes-or-no answer and a magnitude. Here is the one our running example uses:
For large-cap US names, the earnings move implied by the options market is, on average, larger than the move that gets realized.
Notice what makes it well-formed. It names the population (large-cap US names, the biggest companies by market value), the quantity (the implied move, the move size option prices forecast in advance, versus the move actually realized), and the direction of the claim (implied is larger, on average). Compare it to the way trading ideas are usually spoken: "earnings options are overpriced" names no population and no measurement, and "AAPL looks expensive into earnings" is a claim about one name, which the last lesson taught you to walk away from.
Why must the sentence come first? Because if you look at the data before committing to the claim, the data will happily suggest a claim it already supports. You will notice a pattern, convince yourself you suspected it all along, and then "confirm" it on the same numbers that generated it. That is how overfitting starts, tailoring a claim to the accidents of one sample, and no amount of statistical sophistication later can undo it. Writing the hypothesis down first is the cheapest honesty mechanism in all of research.
Step 2: Data
Get the event or series history, point-in-time.
For the earnings study, this means every earnings event in the population, each carrying two numbers: the move the options market implied going in, and the move the stock actually realized after. Assembling that yourself from raw option chains is weeks of work, which is exactly why event-history datasets exist; you will pull this one with a single API call when we get there.
The phrase doing the quiet work in this step is point-in-time: every record must reflect what was knowable at that moment, not what was revised, corrected, or backfilled later. A dataset that shows you Tuesday's numbers polished with Wednesday's knowledge will make any strategy look brilliant, because the strategy is accidentally reading the future. Point-in-time discipline, and the ways datasets quietly violate it, gets a full treatment in the Research Methods module. For now, hold the rule: the data for a study must be the data as it existed on the day.
Step 3: Measurement
Compute the few numbers that answer the claim.
Few is the operative word. The temptation after loading a dataset is to produce forty charts. Resist it: the hypothesis was one sentence, so answering it takes only a handful of numbers. For the earnings study, three suffice.
The event count tells you whether you have evidence or an anecdote. Eight events is a story; several hundred is a distribution.
The mean gap between implied and realized is the headline: on average, was the implied move an overestimate, and by how much?
The hit rate, the fraction of events where implied exceeded realized, tells you the shape. A positive mean driven by 55 percent of events is a very different animal from one driven by 90 percent.
Anything beyond what the hypothesis needs is decoration at best, and at worst it is Step 1 running in reverse: numbers hunting for a claim.
Step 4: Attack
Try to kill your own result.
Suppose the mean gap comes back positive. The amateur move is to celebrate. The professional move is to turn adversarial, because several boring explanations produce the same headline number.
Check the tails: a mean can be positive while rare 25 percent moves destroy anyone actually trading the event. Check the regimes: split the sample by year, because an effect that only existed in one volatile year is not an edge, it is a memory. Check the sample size: means from small samples wobble, and a decimal point does not make forty events a distribution. Check the costs: spreads (the gap between the price you can buy at and the price you can sell at), fees, and, for short strategies, borrow. You saw in What Counts as an Edge: Mechanism-First Thinking how a 4 percent drift can be fully consumed by carry.
If the result survives the attack, you have not found a money printer. You have found something better: a structural baseline you actually understand, measured from real events, that more specific strategies can be built on. And if the result dies under attack, the loop still paid for itself, because it killed the idea before the market charged you to learn the same thing.
The attack step is important enough that it gets its own module: Research Methods is essentially this step expanded into a permanent checklist. The whole course's epistemology lives there.
Why the Order Is the Point
Each step exists to protect the one after it.
Hypothesis before data keeps you from fitting the claim to the sample. Data before measurement forces you to confirm the records are point-in-time before trusting anything computed from them. Measurement before attack gives the attack a specific target instead of a vibe. Run the steps out of order and each protection fails in a predictable way: data first becomes curve fitting, measurement first becomes fishing, and skipping the attack becomes a live trading account performing the attack for you, at much worse prices.
The next module puts tools in your hands: Python, a database, version control, and your first live API calls. From then on, every study you meet will be introduced the same way: here is the hypothesis, here is the data, here is the measurement, now attack it.
Knowledge check
3 questions