13.08.2022 Views

advanced-algorithmic-trading

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

Chapter 16

Supervised Learning

Supervised Learning is the most widely utilised form of machine learning, since it performs two

very general learning tasks across a multitude of domains. These tasks are classification and

regression.

Classification problems involve estimating membership of a set of features into a particular

categorical group. One example might be determining whether a patient is likely to possess

a disease or not (binary class membership) from MRI scan data. A commonly-cited financial

example is predicting whether an asset will rise or fall in price the following day, based on N

days of asset price history.

Regression problems involve estimating a real-valued response from a set of features. One

example would be estimating the sales volume of a product based on separate budget allocations

to different forms of advertising, such as TV, radio or internet ads. A financial example would

be to predict the actual value of an asset in the following day from its past price history, not just

whether it has risen or fallen.

In this book techniques from both aspects of supervised learning will be considered.

16.1 Mathematical Framework

There are two approaches used to provide a mathematical formalism to machine learning.

The first approach formulates the problem in terms of function estimation. The "true"

response variable y (categorical or real-valued) is modelled as a function f = f(x). The predictors

(or features) x ∈ R p is a p-dimensional vector containing feature measurements. In addition to

f there is additional noise ɛ containing information not available within x. ɛ is often assumed to

be normally distributed with mean zero and variance σ 2 :

y = f(x) + ɛ (16.1)

The goal of machine learning under a function estimation approach is to make an estimate

of y, denoted by ŷ by attempting to find a function ˆf which "best" approximates f. Once this ˆf

has been obtained it is straightforward to estimate any new ŷ given a new predictor vector x test .

In some instances however it is harder to choose between values when classifying or regressing,

as there may be some ambiguity from the feature vectors. This motivates the second approach

225

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!