What a trading model actually is
A trading model is a rule that turns an observation into a decision, applied the same way every time. That single property — the same inputs always produce the same call — is what makes a model testable, and testability is what separates it from a forecast you simply have to trust.
The definition that does the work
Strip away the software and the jargon and a trading model is a mapping: it takes a set of inputs — price relative to a level, time, volatility, whatever the rule names — and returns a decision, with the entry, the exit and the size all specified. The defining feature is that the mapping is fixed. Feed it the same situation twice and it must return the same call twice. A discretionary trader is a mapping too, but an inconsistent one: the same chart on a calm Tuesday and a frightening Friday can produce opposite calls. The model removes the mood.
Why mean reversion is the engine of so many models
A great many trading models rest on one durable observation: a price that has stretched unusually far from a typical level tends, often enough to be worth trading, to drift back toward it. A mean-reversion model takes a position on that stretch and waits for the snap-back. It is not a claim to foresee where price goes next; it stakes a position on a recurring pattern, played out over a long run of trades in which the losing ones are tallied just as fully as the winners. The tested models this site uses are all mean-reversion models — they differ from one another only in how long they hold the position.
The same rule on four different clocks
The clock a model trades on is a design choice, not a different idea. The same mean-reversion logic can be expressed on a minutes-long clock, a clock of hours, a clock of weeks, or a long horizon — and the four tested models here do exactly that. Holding for a session is a different discipline from holding for a month, so each clock gets its own model and its own measured record, rather than one model stretched across timeframes it was never scored on.
A worked example: the logic, start to finish
It helps to watch the rule run once, in the abstract, with no instrument named. The schematic below is the whole idea in one picture: a price drifts well below the middle of its own recent range, the model treats that stretch as its only input, and every level is fixed before a position exists.
Suppose a price has fallen for several sessions and now sits an unusual distance below its typical level — further than it has usually strayed before turning. The model does four things in one motion, all decided in advance, none of them a judgement call once the position is open:
stop at 97.00 — a 3.00 point room, past where a further fall would say this is a new trend, not a stretch
target at 104.50 — back near the typical level the price reverted from
risk:reward 3.0 risked to make 4.5 — a 1.5:1 payoff if the snap-back completes
What invalidates the trade is simple and pre-committed: price trading through 97.00 says the stretch was the start of something, not a deviation from a mean, and the model is out. What a realistic outcome distribution looks like is not “this trade wins” — it is that across many such setups the model expects to be wrong a meaningful share of the time, with the winners' reach to 104.50 outweighing the losers' stops at 97.00 often enough to pay. One trade tells you nothing; the distribution is the model.
Notice what the trader never does: decide anything after the position is open. Every field was fixed before the trade existed, which is exactly why the closed record becomes a clean test of the rule rather than a memory coloured by how it felt. The same walk-through runs on any of the four clocks — only the holding window changes.
Sizing the trade: the arithmetic that keeps you in the game
A model is not only an entry rule; it is a size rule. The standard discipline is to risk a small fixed share of the account per decision, so that no single loser can do outsized damage. The maths is worth doing once by hand, because it is the same on every trade.
risk per trade 1.0% → 200 at stake on this decision
stop distance entry 100.00 − stop 97.00 = 3.00 per unit
position size 200 ÷ 3.00 = 66 units (round down, never up)
if stopped 66 × 3.00 = −198 ≈ the 1% you budgeted
The position size falls straight out of the stop distance: a wider stop means fewer units for the same 1% risk, a tighter stop means more. The account percentage, not a gut feeling about conviction, sets the size. That is the whole point — the loss is decided before the trade, not discovered after it.
How a conviction grade maps to size
The tested models add one more layer: a published A-to-D grade marking how strong the setup is relative to that model's own measured returns. A reader who cannot take every call uses the grade to decide where to lean. A common, conservative mapping is to scale the risk budget with the grade — for instance treating an A as the full 1% above, a B as around three-quarters of it, a C as half, and a D as a token or a pass — so that conviction changes the size of the bet, never the discipline of having a stop. The grade is locked in before the result lands and cannot be nudged upward once a call goes green, which is what makes it safe to size on.
| Model | Holding clock | Grade-A bar (per trade) |
|---|---|---|
| Day Trade | opens and closes inside the same session, a 0 to 60 minute window | 0.70% avg / trade |
| Multi Hour | runs from half a session to roughly two sessions | 4.50% avg / trade |
| Swing Trade | carries a position for about 7 to 28 days | 6.00% avg / trade |
| Investing | is held across a long horizon | long-horizon |
An A sits in the top band of a model's own measured return spread; D is the lowest grade still published. Because the bar is set per clock, an A on a same-session call (around 0.70% a trade) and an A on a multi-week Swing call (around 6.00%) both read as “top-band for this horizon” rather than one absolute number stretched across very different holding times. There is no E grade — it was retired so the four-step scale keeps its meaning.
What a bad version of this looks like
The fastest way to understand a good model is to see the fragile, dishonest versions it is meant to replace. Every one of these is a way of keeping the appearance of a model while throwing away the property that made it worth anything.
- Vague entries. “Buy when it looks oversold” is a mood, not a rule. If two people reading it would act differently, it cannot be tested, and an untestable model can score almost any outcome as a success after the fact.
- A moving stop. Widening the stop once price approaches it — “giving it room” — turns a 1% risk into an open-ended one and means the model you are running is no longer the model you tested.
- No size rule. Sizing by how confident you feel, rather than by the stop distance and a fixed account percentage, is how a single bad trade erases a month of good ones.
- An overfitted backtest. A curve tuned until it fits one stretch of history perfectly has usually modelled the noise, not an edge; it looks spectacular on the past and breaks on the first live month.
- A highlight-reel record. Showing the winners and quietly dropping the losers is not measurement; it is marketing. A record without its losers and its drawdown has measured the flattering half and skipped the half that decides whether the edge survives.
How do you actually know a strategy works?
The honest question to ask of any model — yours or someone else's — is not “how good do the numbers look” but “how hard would they be to fake.” The weakest answer is a backtest, because a curve can be tuned until it fits the noise in old data and still fail the moment it meets a live market. A stronger answer is a forward record: a continuous series of calls, each logged before its outcome, with the losers left in. The strongest answer of all is a forward record a stranger can re-check without trusting the person who kept it.
That is exactly the standard the tested models here are held to. Every published call carries an A-to-D conviction grade, and the whole call — its entry, target, stop, grade and signal time — is condensed into a single cryptographic fingerprint that is anchored to a public ledger the instant it goes out. Touch any field afterward and the fingerprint no longer matches the receipt, so a confirmed match proves the call existed in that exact shape before the trade resolved. A record built this way is not a story you are asked to believe; it is a claim you can audit, call by call. Reading such a record line by line is a skill worth practising on any model, tested or your own.
The dividing line, in the end, is testability. A real model can be run against history and forward in time and made to produce a record; a guess produces only a highlight reel of the times it happened to be right. The cleanest proof that a model was a model and not a story told afterward is that each call was committed before its outcome was known. That is why the tested example here — the #1-ranked provider's four mean-reversion models — anchors every call on a public ledger at publication: it turns “decided in advance” into something a stranger can confirm rather than something you simply have to trust.