The are a lot of misconceptions regarding neural networks. This article, and subsequent ones on the topic will present the major ones one need to take into account.(1)
A very common misconception is that NNs require big data in order to work. This is not correct. They have advantages when used with small sets of data as well, depending on the problem.
Neural networks is thought to require lots of data for training, the more the better. However, this is not correct as it depends on the problem at hand. The crucial things is the learning strategy, which determines the amount of data required.
Neural networks can use one of three learning strategies:
- supervised learning: requires at least two data sets, a training set which consists of inputs with the expected output, and a testing set which consists of inputs without the expected output. Both of these data sets must consist of labelled data, i.e. data patterns for which the target is known upfront.
- unsupervised learning: typically used to discover hidden structures (such as hidden Markov chains) in unlabelled data. They behave in a similar way to clustering algorithms.
- reinforcement learning: based on the simple premise of rewarding neural networks for good behaviours and punishing them for bad behaviours.
Because unsupervised and reinforcement learning strategies do not require that data be labelled they can be applied to under-formulated problems where the correct output is not known.
As we have already dealt with supervised learning in a previous article, we will describe the other two below.
One of the most popular unsupervised neural network architectures is the Self-organising Map (also known as the Kohonen Map).
Self-organising Maps are essentially a multi-dimensional scaling technique which construct an approximation of the probability density function of some underlying data set, whilst preserving the topological structure of that data set. This is done by mapping input vectors in the data set to weight vectors, (neurons) in the feature map.
Preserving the topological structure simply means that if two input vectors are close together in set, then the output neurons to which those input vectors map in will also be close together. An example Kohonen map is show below.
Reinforcement learning strategies consist of three components.
- A policy which specifies how the neural network will make decisions, e.g., using technical and fundamental indicators.
- A reward function which distinguishes good from bad, e.g., making vs. losing money.
- And a value function which specifies the long term goal.
In the context of financial markets (and game playing) reinforcement learning strategies are particularly useful because the neural network learns to optimise a particular quantity such as an appropriate measure of risk (e.g., for adjusted return).
In the financial realm we may want to use a reinforcement learning NN to determine whether to sell or buy stocks. An example is shown below.'
In essence, the goal is to make money: thus, when this succeeds, we reinforce the weights that led to that result. In the opposite when not. This is similar to backpropagation, but necessitates a connection from the real world to the network, and is thus not totally internal to the network architecture.
(1) The inspiration for the misconceptions is adapted from an article by Stuart Reid from 8 May 2014 available at http://www.turingfinance.com/misconceptions-about-neural-networks/.