How it works

From market data to a logged result.

Signals are generated by machine-learning software developed by our broker partners. Its design, data and parameters are covered by a non-disclosure agreement. The software does not trade anyone's account: you place your own trades, in your own account.

A calculation is not knowledge of the future.
Every idea needs a level where it is declared wrong. That level is the stop-loss.
A signal is not an order. You decide whether to take it and how much to risk.

01 · Signal anatomy

How a signal is born.

Click through the eight steps. Each one shows what happens, what you decide, what can go wrong and what stays private under the NDA.

01 · Market data

Illustrative example

Signal software starts from market prices: gold, silver and the major currency pairs. Gaps, bad ticks and session changes have to be handled before anything is calculated.

What you decide: Nothing yet. This is the raw material.
What can go wrong: A data error can look like a real move. Garbage in, garbage out.
What we do not disclose: Data sources and cleaning rules (NDA).

02 · Math lab

Math lab

Interactive models that explain the ideas behind the panels. Everything here runs on synthetic data and is for teaching only.

Mean-reversion lab

Synthetic data · teaching onlyOU PROCESS

κ · speed of return0.08

θ · long-run mean0.0

σ · volatility1.0

spread

Half-life = ln2 / κ8.7 steps

Stationary σ2.55

The dots mark |z| > 2, purely as a teaching device: the spread is unusually far from its mean. That is a question, not an order. On analysis panels of this kind, β is the hedge ratio, Z-score the distance from the mean in standard deviations, half-life the typical time to close half the gap, θ, κ, σ the parameters above, and the ADF p-value a test of whether the series mean-reverts at all.

Pairs idea: gold vs silver

Synthetic data · teaching onlyRATIO / SPREAD

Regime break at step200

market A (gold-like)market B (silver-like)

naive spread z-score

Two related markets usually move together, so the spread between them (B − β·A) tends to come back to its mean. After the break, the relationship drifts and the naive model keeps saying "too far, it will come back" while the gap keeps growing. That is why every idea needs a level where it is declared wrong.

Kalman filter toy

Synthetic data · teaching onlyADAPTIVE β

Adaptivity (process noise, log10 q)1e-4

true β (hidden)Kalman estimatefixed estimate

The hedge ratio between two markets drifts. A fixed estimate from the first weeks goes stale; an adaptive filter follows the drift. More adaptivity tracks faster but is noisier.

Structural-break demo

Synthetic data · teaching onlyCALM → VOLATILE

Break point150

Add one big jump in the calm part

price

rolling volatility

A real regime change shows up as a lasting shift in volatility. A single jump lifts rolling volatility for a short while and then fades: one jump is not proof of a regime change.

Agent-based market toy

Synthetic data · teaching onlyEDUCATIONAL MODEL OF PARTICIPANTS

Follow rises40

Buy after drops40

Reduce when volatile20

run 1run 2run 3

Three simple types of participant, three random runs with the same mix. When trend followers dominate, moves extend; when contrarians dominate, prices chop around. Small changes in the mix give very different paths: this is scenario sensitivity, and the reason a model is never a crystal ball.

Itô's lemma, in plain English

Background maths

df (t, X_{t}) = (\frac{\partial f}{\partial t} + μ \frac{\partial f}{\partial x} + \frac{1}{2} σ^{2} \frac{\partial ^{2} f}{\partial x ^{2}}) d t + σ \frac{\partial f}{\partial x} d W_{t}

Prices move with a random component. Itô's lemma is the rule for how a function of such a random price changes over time. The extra ½σ² term is the surprise: randomness itself changes the average path. This is the maths behind option pricing and volatility models, and quantitative software often reports it next to statistics such as skewness, kurtosis, the Hurst exponent and ACF(1). It is background maths that describes how uncertainty behaves. It does not predict where a price will go.

03 · Glossary

Statistics glossary

Common analysis metrics, one line each.

Skewness

Synthetic

How lopsided the distribution of returns is.

Why a trader cares · A strategy with many small wins and rare big losses has negative skew. The average hides that tail.

Excess kurtosis

Synthetic

How fat the tails are compared with a normal distribution.

Why a trader cares · Fat tails mean extreme moves happen more often than a bell curve suggests. Stops must allow for them.

Hurst exponent

Synthetic

Below 0.5 a series tends to revert, above 0.5 it tends to trend, 0.5 is a random walk.

Why a trader cares · It tells you whether a mean-reversion or a trend idea fits the market right now.

ACF(1)

Synthetic

Correlation of a value with the previous one.

Why a trader cares · Negative values hint at reversal from one step to the next, positive values at momentum.

ADF p-value

Synthetic

A statistical test of whether a series is stationary (has a stable mean).

Why a trader cares · A low p-value supports a mean-reversion idea. A high one says the spread may simply wander off.

Half-life

Synthetic

Typical time for a gap from the mean to close by half: ln2 / κ.

Why a trader cares · It sets expectations for how long a trade might need, and when it is taking too long.

04 · Video

The explainer, chapter by chapter.

Video coming soon

The 10-chapter explainer (about 7 minutes) will appear here with captions and a transcript.

Presented by Yanis, the trader behind the record.