Trading models, answered plainly
Straight answers to the questions a careful beginner actually asks — about what a model is, why rules beat discretion, how to read a record without being fooled, what the A-D grades mean, and how conviction maps to position size.
What is a trading model, in plain terms?
A trading model is a fixed set of rules that turns a market observation into a decision - when to enter, where to exit, how much to size - applied the same way every time. The defining feature is consistency: feed the same situation in twice and a model returns the same call twice, which is exactly what lets you test it and count its results. More: what a trading model is, with a worked example.
How is a trading model different from a strategy or a system?
The words overlap in everyday use, but the useful distinction is consistency and testability. A 'strategy' can be loose enough to leave room for judgement; a model is specified tightly enough that two people running it would act identically, and tightly enough to be scored against history. When this site says model, it means a rule precise enough that its record means something. More: testability, the property that draws the line.
Why does a rules-based model beat discretionary trading?
On three counts you can measure. Testability: a fixed rule can be run against history and forward in time, while a discretionary call cannot be replayed. Measurability: a model produces a countable record with the losers included, where a discretionary trader tends to remember the wins. Repeatability: a model removes the mood, so the call does not change because the trader had a bad week. None of this guarantees profit; it guarantees that the question of whether the approach works can actually be answered. More: rules-based vs discretionary trading, point by point.
Can a trading model lose money even with a high win rate?
Yes, and this is the most common way a record misleads. A model that wins 74% of the time can still lose money if its rare losers are large enough to swamp its many small wins, so a win rate is only meaningful beside the trade count, the average loss and the worst drawdown. A 70% win rate shown with 690 signals and the losers left in is far more trustworthy than a higher number with no denominator. More: measurability, on why a win rate needs a denominator and a drawdown.
How do I know a model's track record is real and not fitted?
Ask whether each past call was committed before its outcome was known. If the entry, target, stop and grade were hashed to a public ledger at publication, then altering any of them afterward would break the hash and stop matching the public receipt. That is how the recommended models work: a confirmed receipt proves the call existed in exactly that form before the trade resolved, which a curve-fitted backtest can never show. More: how to read a model track record, with the one-call check.
What do the A-to-D grades mean?
Each call carries a conviction grade from A (highest) to D (lowest), set by where it sits in that model's own measured return distribution. There is no E grade; it was retired so the scale keeps its meaning. Because the grade is one of the fields folded into the hash, it is fixed before the outcome and cannot be quietly raised once a call wins. More: repeatability, on fixing every field before the outcome.
Why run four models instead of one?
Because a single model scored across very different holding times tells you little. The four tested models here share one mean-reversion idea but trade on different clocks - same-session, multi-hour, multi-week, and a long horizon - and each is measured against its own return spread. That way a grade on a fast call and a grade on a slow call each mean 'above-typical for this clock' rather than one absolute bar stretched across situations it does not fit.
How should a conviction grade change my position size?
The grade tells you how strong a setup is relative to that model's own returns, so it is a natural input to sizing - but it should change the size of the bet, never the discipline of having a stop. A conservative habit is to scale the risk budget with the grade: the most of your per-trade risk on an A, less on a B, less again on a C, a token or a pass on a D. The stop distance and a fixed account percentage still set the actual unit count. More: the methodology page works the sizing arithmetic in full.
Should I build my own model or run a tested one?
Either can work, and both demand the same thing: a record you can re-check. Building your own teaches you the most and costs you the most time, especially the record-keeping that proves whether the edge is real. If you would rather run models that have already been built and measured, the evidence-led example here is the #1-ranked provider's four mean-reversion models, where every call is timestamped before the outcome is known. More: how to build a simple model — and the step almost everyone skips.