Neural Network Wizard

Neural Networks for Traders

You don't need any maths to use the Wizard well — but a handful of plain-English ideas will make every setting in this manual click into place. This page is that foundation. Read it once and the rest of the guide becomes a series of obvious choices rather than mysterious dials.

What a neural network actually is

A neural network is a flexible pattern-finder. You show it many examples of some inputs (say, today's RSI, the slope of an EMA, and yesterday's return) paired with the answer you wish you'd known (say, the return five bars later), and it gradually works out how the inputs relate to the answer.

Think of a moving-average crossover: you decided the rule yourself — "go long when the fast MA crosses the slow MA". A neural network is the same idea turned inside out. Instead of you picking the rule, the network discovers the rule from history, including subtle, non-linear combinations of indicators that would be hopeless to write by hand. It is just a very flexible way to blend several indicators into one number.

Note

The network is built from layers of simple units called neurons: an input layer (one slot per input you feed it), one or more hidden layers that do the pattern-finding, and an output layer that gives the prediction. More neurons and layers let it learn more complicated relationships — and, as you'll see, make it easier to fool. The Architecture tab is where you size all this.

Training vs. prediction

Working with a network is always two separate phases — keep them straight and everything else follows:

  1. Training. The network studies your historical bars and slowly tunes thousands of internal numbers (its weights) until its guesses match the real answers as closely as it can. You do this once, and the result is saved to a .net file.
  2. Prediction. You load that trained network and apply it to new bars to get a forecast. No more learning happens — it simply applies what it already learned.

That split is exactly what the Wizard generates for a Standard network: a Training Formula and an Indicator Formula. The How It Works page walks through the loop end to end.

The one mistake that matters: overfitting

If you remember a single thing from this manual, make it this. The great danger with any neural network is overfitting — the network memorises the quirks and random noise of your training data instead of learning a real, repeatable pattern. An overfit network looks brilliant on the bars it trained on and falls apart on new bars, which is the only thing that matters for trading.

The skill you actually care about is generalisation: doing well on data it has never seen. The whole craft of building a network for the markets is a tug-of-war between giving it enough power to find the pattern and not so much that it starts memorising. Financial data is mostly noise with a faint signal, so this is easier to get wrong here than almost anywhere else.

Tip

Three habits keep overfitting in check, and they cost you nothing: more data (longer history, and training across several symbols rather than one), simpler models (fewer neurons and layers — start small), and holding back a test set so you can actually see whether the network generalises. The entire Accuracy tab exists to help you win this fight.

Garbage in, garbage out

A network is not magic. It can only find a relationship between your inputs and your target if one genuinely exists, so choosing good inputs and a sensible target is the most important work you do. The network just finds the best combination of whatever you hand it.

One consequence matters to every trader. A network works best with data in a small, steady range, so raw price is a poor input — a stock that traded at $5 a decade ago and $300 today never sits still. Feed it a rate of change (the percent change of price) instead, and the numbers stay in a roughly constant range across the whole history. That single habit — the Use Percent Change option — makes predictions far more reliable and lets one network work across many symbols.

A prediction is only the start of a system

The Wizard gives you a prediction — a line on a chart. That is the beginning of a trading system, not the end of one. You still have to decide where to enter and exit, account for spread, commission and slippage, size your positions and manage risk. A forecast that is right 55% of the time can still lose money if the trade plan around it is poor.

Be realistic, too: predicting the next return from scratch is genuinely hard, and the signal is usually thin. A gentler, often more rewarding use is to keep a system you already trust and let a network filter its signals — keeping the entries that look like winners and dropping the likely losers. It's an easier question to ask, and it tends to answer better.

Note

Neural networks are a useful tool, not a holy grail. Treat each prediction as one more input to your own judgement and risk management, validate it honestly on data the network never trained on, and you'll get the most out of the Wizard.

Where to go next