The Laws of Trading by Agustin Lebron

Author:Agustin Lebron
Language: eng
Format: epub
ISBN: 9781119574200
Publisher: Wiley
Published: 2019-06-18T00:00:00+00:00

Characteristics of Good Models

So, what do good models look like? According to the rule, good models are well-adapted to the edges they express. This adaptation guides the consideration and evaluation of models along a few different dimensions.

Robustness

The robustness of a model refers to its ability to maintain performance when you consider both aspects of the model and the world in which it lives and interacts:

If historical data is used to train the model, does it perform similarly when different historical periods are used for training? If not, then the model probably isn't expressing a very reliable or robust edge.

In particular, how does the model perform on a separate piece of historical data once trained? This is known as out-of-sample testing. Again, a model that performs well on data it's seen and poorly on data it hasn't seen is one whose edge is unlikely to be reliable.

Once a model's parameters are determined, how does its performance change when these parameters are tweaked by small amounts? If a small tweak of a particular parameter significantly degrades the performance of the model, it's not robust.

How sensitive is the model to the data sources it uses? If the data has outages, or starts to exhibit weird or stale information, will the model's performance degrade gracefully? Or will it start to incinerate money? This is a common failure mode for trading firms.

How usable is the model in the real world? Does it require a lot of moving parts, is its architecture reliable or prone to breakage? An excellent predictive model isn't much good if the systems on which it runs are always going down with some fault or other.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.