Machine learning models for stock prediction

A trend line uses one clue: the day number. A moving average uses one clue: the recent average. Machine learning is greedy in the best way: it takes many clues at once and learns, from the data itself, how much to trust each one. Here you will build the clues, split the data properly, and train a real model right in your browser.

From one clue to many: features

Think about predicting whether your friend will be late. You do not use one fact, you blend several: how they slept, the traffic, the weather, their track record. Each fact is a feature. To predict NCB's price for a day, we build features from the days just before it:

Every feature is built only from days before the one we predict. That rule is sacred: if a feature ever peeks at the answer, the model looks brilliant in practice and fails in real life.

The golden rule: train and test

Here is a trap. If you let a model study all the data and then grade it on that same data, of course it scores well, it has seen the answers. That tells you nothing about the future. So we split the days in time:

Splitting by time, not at random, matters for prices: you must always train on the past and test on the future, never the reverse. The test error is the only number you should believe.

The model itself

Our model is a linear regression, the number-predicting cousin of the win-or-lose model from the football course. It gives each feature a weight and adds them up:

"Training" just means the computer searches for the weights \(w\) that make the error on the training days as small as possible, the same error you measured by hand in Lesson 3. Then it uses those weights on the test days it never saw.

The truth about markets

Notice something in the lab: even a good model only just edges out "tomorrow equals today," and on calm stretches it can lose. That is not a bug. Prices are close to random in the short term because thousands of smart people are already trading on every scrap of news, so the easy patterns get used up almost instantly. This idea has a name, the efficient market, and it is why nobody can reliably predict tomorrow's exact price.

The three machine learning models the pros use

The browser lab used a linear regression, the gentlest machine learning model. When analysts want more power on real stock data, three models do most of the heavy lifting.

• Random forest grows hundreds of decision trees on random slices of the data and averages their votes. It rarely overreacts to one weird day, which makes it a safe first pick.
• Gradient boosting (the famous version is XGBoost) builds trees in a chain, each new tree focused on the mistakes of the ones before it. On tabular features like ours it is often the single strongest model, and a favourite in real trading competitions.
• LSTM (Long Short-Term Memory) is a neural network with a memory, designed for sequences like prices, text, and music. It can catch longer patterns a single tree would miss, but it is hungry for data and computing power.

Notice the pattern: all three build on ideas from the football course, decision trees and neural networks, just scaled up and pointed at prices.

Try each model yourself

Each machine learning model now has its own hands-on deep dive: a plain-words explanation, a game you can play, and a live chart on real NCB data. Grow a forest, chain a booster, and test a memory cell.

Random forestGrow trees one by one and watch their votes average out.Open the interactive → Gradient boostingAdd correcting trees that shrink the leftover errors.Open the interactive → LSTM neural netPlay with a forget gate and test the model's memory.Open the interactive →

Read the moving average

Here is AI Boss Inc again, this time with its 20-day moving average drawn in gold on top of the daily price.