Thread by @pedma7

"Smoothing Volatile Alpha"

We all read concepts that we don't fully understand what they mean or their implications.

That's fine, we don't know everything.

I try to understand by writing.

Let's go over this concept and simplify what's discussed below.

https://twitter.com/therobotjames/status/1560362519713107968?s=20

I want to start by saying that I am exploring the concepts in this tweet and by no means am I going to represent them in perfection.

I welcome anyone to correct me in the way I am representing the thoughts here shared.

Let's imagine we are trying to predict tomorrow's weather.

We have a feature or set of features that give us a prediction.

However these features are very "jumpy" and change their mind a lot.

This is the volatile alpha that is mentioned in the conversation.

Now because our weather detection system is so jumpy, we try to average out its predictions over a few days to make it more stable.

This is what's meant by "smoothing" the alpha.

However, as always, there's a catch. The more you average the signal, the less accurate it becomes.

That loss in accuracy is similar to the "decay in IC" that is talked about here.

IC = Information Coefficient.

However we notice that when we average out those predictions, they tend to follow a certain pattern or trend.

This is what is meant by autocorrelation.

Autocorrelation, in the context of our weather prediction, it's like noticing that it has been sunny for the past few days, and due to that it's likely to remain sunny for the remaining of the days.

Let's now get into @macrocephalopod answers.

Continuing with our weather analogy, let's say you're planning a picnic tomorrow.

You care a lot about tomorrow's weather.

But if you're planning a week long camping trip, you care more about overall trend.

https://twitter.com/macrocephalopod/status/1560363204429201408?s=20

If you're looking at the weather for a week-long trip, don't average out the predictions for a whole month.

Maybe average out just a week or so.

This is the idea behind "smoothing based on holding period".

https://twitter.com/macrocephalopod/status/1560363618864087042?s=20

And instead of just relying on the "jumpy" set of features we have, maybe use other tools or methods to decide when to have the picnic or camping trip.

Like trading, every time you go on that picnic or camping trip you will incur in costs.

https://twitter.com/macrocephalopod/status/1560364099422371843?s=20

This decision of going on trip or not is the analogy to the "turnover" in trading when we buy and sell stocks.

Every time you decide to go out, based on the tool's prediction, there's a risk or cost, like getting caught in rain.

This is the transaction cost.

If you go out too often, without considering the risks, you will get wet more often.

When planning those outdoor activities, you should account for those risks.

If we have a specific type risk (quadratic transaction cost), with a factor "q" it will affect how often you go out.

The higher the value of "q", the less often you'll want to go out because the risk becomes higher.

The formula "1/(1+q)" is a way of saying that the more risk factor "q" increases, the less often you'll decide to go out.

When "q=0", there's no risk, so you'd go out as often as the jumpy weather tool suggests.

Bust as "q" gets bigger, you'll decide to stay in more often.

Accounting for the risk "q" has a similar effect to smoothing, because increased cost will "punish" outcomes of jumpy tools.

In essence, the conversation is about how to best use a prediction tool that's a bit jumpy and how to make decisions based on its predictions.

The key takeaway is the importance of considering the broader context of what's trying to be achieved and understanding inherent risks.