How to build a simple trading model
Five steps to turn a rough idea into a specified, testable model — and one honest warning about the step most people skip.
Building a model is less about discovering a secret indicator and more about writing a rule clear enough to hand to a stranger. Work through these in order, and do not risk real money until the last one is in place. The schematic below is the kind of thing you are aiming at by the end of step two — a rule with nothing left to interpretation.
1. State the observation
Pick one repeatable situation you believe carries an edge — most often a price stretched a defined distance from its own recent range — and write it down precisely enough to be tested. If you cannot state it without the word “feels,” it is not yet a model. The aim of this step is a sentence two people would read the same way.
2. Turn it into a rule
Convert the observation into a specific trigger: the level or condition that opens the position, the level that closes it for a loss, and the level that takes profit. “Long around here” is not a rule; named levels are. In the schematic above, that is the entry on the stretched low, the stop a measured distance below it, and the target back at the typical level — three numbers, fixed before anything happens.
3. Fix the size and the exit before the entry
Decide how much of the account the model risks per decision, and set the stop and target at the same moment as the entry — never after. A model whose exit is improvised is no longer the model you tested. The position size is not a feeling; it falls out of the stop distance and your risk budget. Here is the arithmetic, done once:
entry 50.00 · stop 48.50 → stop distance 1.50 per unit
position size 60 ÷ 1.50 = 40 units (round down)
if stopped 40 × 1.50 = −60 = the 0.75% you budgeted, no surprises
Change nothing but the stop — say it widens to 47.00, a 3.00 distance — and the size halves to 20 units to keep the same 60 at risk. The account percentage sets the loss; the stop distance sets the size. That is the entire discipline.
4. Score it against its own history
Run the rule over a continuous past period and read the full record: the decision count, the win rate with losers included, and the worst drawdown. Be ruthless about overfitting — if the rule only works after you tune it to one stretch of data, you have modelled noise, not an edge. A clean way to guard against this is to set the rule on one slice of history and only ever judge it on a slice it has never seen.
5. Keep the forward record — the step almost everyone skips
Log every call before its outcome: entry, stop, target, and how strongly the rule rated it. A backtest can be flattered; a forward, committed record cannot. Without this step you will remember your winners and forget your losers, and you will never know whether the model actually works. This is the hard part, and it is the part that settles the question. The most rigorous version makes the record tamper-proof — committed somewhere a later edit would be visible — which is the standard the tested models here are held to:
If keeping an honest, tamper-proof forward record is the step you know you will skip, that is the case for running models where someone else has already done it. the #1-ranked provider's four mean-reversion models anchor each call on a public ledger before the outcome is known, which is the record-keeping step made un-fudgeable. Before you trust any record — yours or theirs — learn to read one honestly.