Framing Machine Learning as an Allocation Problem

Summary

The document outlines a supervised learning setup for choosing positions in a single instrument. At each point in time, past prices are summarized as features such as momentum, volatility, and other technical indicators; fundamental information could also be included. A model, such as a neural network or random forest, would use those inputs to produce a position for a future period.

Its central issue is how to define the training target. Future price change is observable, but predicting that change does not directly specify how much capital or risk to allocate. The text raises the need to connect forecasts to position decisions, but it does not propose an objective, allocation rule, or solution. It presents a useful problem framing rather than a tested method, and gives no data, empirical results, or discussion of costs, risk constraints, or validation.

Key ideas

• Price history can be compressed into features such as momentum and volatility.
• Fundamental variables can be added to the model inputs when available.
• A model could map current features to a position for a future period.
• Predicting returns alone does not determine the appropriate allocation.
• The document poses the target-design problem but does not solve it.

Tags

# How to use machine learning to generate optimal allocations for an instrument?


# How to use machine learning to generate optimal allocations for an instrument?












What is the idea behind using Machine Learning in finance? Let's assume that we have just one instrument given by its prices. At a given moment of time, we can "compress" the available history of prices into a vector of features (for example, momentum, volatility, some other technical indicators). In addition to that we can extend the features vector by some fundamental factors (like company parameters at a given moment of time).

Now we want to have a machine learning model (for example a neural network or random forest) that take the extended vector of features as input and generate the optimal position for the given future period of time.

The problem is that we do not have target to train the model. Yes, we do have the price change for the given period of time but should we try to predict it? Or, in other words, let's assume that we do have a model that predicts the observed price changes with some accuracy, how do we transform these predictions into allocations?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.