AI and machine learning: supervised, unsupervised, reinforcement, deep

What are AI and machine learning?

Artificial intelligence (AI) is computer systems performing tasks that traditionally needed human intelligence, such as recognizing patterns and making decisions. In some limited settings, AI systems now match or outperform people. An early form was the expert system : hand-coded “if–then” rules written by human experts.

Machine learning (ML) takes a different route. Nobody writes the rules. The algorithm is given data and learns how inputs relate to outputs, or how the data are structured. The slogan is find the pattern, apply the pattern. ML is used across finance, from fraud detection and transaction pricing to optimizing trading strategies.

ML needs a lot of data to train on. Big data supplied it, while computing power grew and storage got cheaper. A robo-adviser is one result: it builds a diversified portfolio from a client’s goals and risk tolerance and keeps rebalancing it automatically.

How is the data split?

The algorithm gets inputs and, sometimes, outputs (targets). The data are split into three sets.

Available data

• Training dataset Memorize learn; the largest
• Validation dataset Memorize tune
• Test dataset Memorize judge on unseen data

• The training dataset is where the model learns input–output relationships. It is typically the largest.
• The validation dataset checks those relationships and is used to tune the model.
• The test dataset (also called the testing dataset) judges how well the model predicts on data it has never seen.

A model is judged on the test set, not on how well it explains its own training data. How big each set should be depends on how much data there is. These are rules of thumb.

Dataset	Training	Validation	Test
Standard	60–80%	10–20%	10–20%
Larger datasets	70–80%	10–15%	10–15%
Smaller datasets	60–70%	15–20%	15–20%

Bigger datasets can spare a larger share for training; small ones keep more back to check the model.

For most data, the rows are shuffled before splitting, so each set is representative and any ordering in the file cannot bias the model. Shuffling also helps prevent data leakage : information from outside the training data slipping into what the model learns, which makes test results look better than real-world performance.

Why not shuffle a time series?

In financial time series, order matters: today depends on yesterday. Shuffling breaks that order, creates relationships that do not exist, and can put future data in the training set while earlier data sit in the test set. That is data leakage. Use one of two methods that keep the order.

Splitting a time series

• Random shuffle breaks time order; leaks the future
• Time-based split Memorize earlier trains, later tests
• Rolling-window validation Memorize fixed window, moved forward

• A time-based split uses earlier data to train and later data to validate and test.
• Rolling-window validation trains on a fixed-length window, validates on the period right after it, then moves the window forward and repeats.

Exam tip Validation and test sets should be at least as long as the forecast horizon. To forecast 12 months ahead, test on at least 12 months. (In the example, 18 months is enough.)

A trained model can be used on other datasets only if they are similar enough. Even a model trained on one group of shares may need further training (fine-tuning) before it predicts another group well. Human judgment stays essential: understanding the data, cleaning it, choosing the method and the model structure, handling outliers and volatility shifts, and controlling bias. Small datasets need extra care.

What are the four approaches to ML?

Machine learningfind the pattern, apply the pattern

• Supervised learning Memorize labeled inputs and outputs; predicts or classifies
• Unsupervised learning Memorize unlabeled data; finds structure and groups
• Reinforcement learning Memorize rewards and penalties; trial and error
• Deep learning Memorize neural networks with many layers
  • Supervised
  • Unsupervised
  • Semi-supervised

Approach	Learns from	Finance example	Strengths	Limits
Supervised	labeled inputs and outputs	predict whether a market index will be up, down or flat tomorrow; predict bond downgrades from past bonds with known outcomes	accurate with enough data; well understood	needs lots of labeled data; may not generalize to entirely new situations
Unsupervised	unlabeled data	group 2,000 companies into peer groups by their financial ratios instead of sector labels; flag unusual transactions	finds hidden groups; can reduce the number of variables (dimensionality); good for exploring	results hard to interpret; patterns may not be meaningful; judging results is subjective
Reinforcement	rewards and penalties from interacting with an environment (trial and error)	a trading agent rewarded when its buy, sell or hold decisions raise risk-adjusted return, penalized when they cause losses	suits step-by-step decisions in changing settings; learns from experience	sensitive to a badly designed reward; needs heavy computing and long training
Deep learning	many-layered neural networks; supervised, unsupervised or semi-supervised	read text, recognize objects in satellite images	models complex, non-linear patterns; extracts features from raw data	needs large data and computing; black box; can overfit unless regularized

In supervised learning , the training data carry labels, so the algorithm learns the link between inputs and a known answer, then predicts for new data. In unsupervised learning no answer is provided; the algorithm describes the data and finds groups. In reinforcement learning a program learns by trial and error, acting in an environment and receiving rewards and penalties. Deep learning uses neural networks with many hidden layers to process data in stages, building simple concepts into complex ones. It can be supervised, unsupervised or semi-supervised.

Exam tip Some summaries list only three main types (supervised, unsupervised, deep). Learn all four: reinforcement learning is the one that learns from rewards, with no labels.

Memorize

How can ML go wrong?

Model fit

• Underfitted Memorize too simple; misses real patterns
• Well fitted captures signal; generalizes
• Overfitted Memorize too complex; learns noise; fails on new data

Overfitting happens when the model learns the training data too precisely. It is overtrained: noise is treated as signal, so it discovers false relationships. The model is too complex, scores brilliantly in-sample and predicts poorly on a different dataset. In symbols, an overfit model has a training error far below its test error: $\text{Error}_{train} \ll \text{Error}_{test}$. A model that shows underfitting does the reverse: it treats real patterns as noise, is too simple, and does badly even on its training data.

Exam tip Great training score, poor test score means overfit. Poor everywhere means underfit.

Because nobody programs the rules, ML models can be black box models: their results are hard to explain. The label comes from their complexity and lack of interpretability, not from bias in the data.

Data mining is a subset of data science that searches large datasets for patterns with algorithms, for example ones built on Bayes’ theorem. Mining without real-world context finds patterns that are only noise, so domain expertise must confirm that a pattern makes sense.

Go deeper Overfitting, data mining and p-hacking

Overfitting is the machine-learning cousin of p-hacking in hypothesis testing. Try enough models or patterns on the same history and one will look good by chance. Both fail the same test: the result is not replicated on new data. That is why ML keeps a test set aside.

Go deeper Explainability and model risk

A black-box model can predict well and still be hard to defend. Regulated firms must be able to explain why a decision was made, and an unexplained model can break quietly when markets change. Human oversight of the data, the method and the output remains part of the process. The last lesson of this module covers how regulators and firms handle model risk.

How does a neural network learn?

A neural network is layers of connected nodes (neurons). The input layer has one neuron per feature, for example one per financial ratio. One or more hidden layers transform the data step by step (many hidden layers make the network “deep”). The output layer gives a classification (downgrade or not) or a number (a forecast return). Each connection has a weight. Neural networks date from 1958 and have been used for forecasting and pattern recognition since the early 1990s.

Training a neural networkrepeated over many epochs

• Forward pass predict
• Loss calculation measure the error
• Backpropagation find each weight’s effect
• Gradient descent step against the gradient

Training repeats four steps:

1. Forward pass. Data flow through the layers to a prediction.
2. Loss. The error between the prediction and the actual value is measured.
3. Backpropagation . The loss gradients show how each weight affects the error.
4. Gradient descent . Each weight is moved a little in the direction opposite to the gradient, which lowers the loss.

One full pass over the training data is an epoch . Training runs many epochs until accuracy is good enough.

Generative adversarial networks and variational autoencoders

A generative adversarial network (GAN) pairs a generator network, which makes synthetic data, with a discriminator network, which tries to tell the synthetic data from real data. In finance it is used to create realistic synthetic data for training and for stress-testing strategies. A variational autoencoder (VAE) is an encoder–decoder network that compresses data and reconstructs it. It is used to reduce dimensionality and to detect anomalies.

Method	Purpose	Structure	Type
Backpropagation	train (optimize) a network	one network	supervised
GAN	generate realistic data	generator + discriminator	adversarial
VAE	learn a compact representation	encoder–decoder	probabilistic generative

Generative AI builds on GANs and VAEs; the last lesson of this module covers it.