Thread by @therobotjames

People say a lot of things about trading, and most of it is worthless.

So it's useful to be able to quickly discard ideas.

Let's take an example...

Many in crypto, including @zhusu, will tell you that buying new highs is a good plan.

Is it? Let's have a look...

1/n

It's important to understand that discarding ideas is a lot easier and quicker than verifying ideas.

Your mission is not to do the most perfect simulation of reality from the offset. You'll waste a lot of time doing that.

You want to do very quick data analysis.

2/n

Plenty of time to go deep later.

We:
1. pull daily price data for all FTX spot contracts
3. for each asset for each day, calculate the 20d high
4. calculate the distance in days from the 20d high
5. calculate next day log returns

3/n

Then we summarize all our data.

We group by the number of the days from the high, and plot the mean next day return for each group.

If our friends are right, we expect to see higher next day returns the closer we are to a high... and lower the further away.

And we do.

4/n

So at least on average, we haven't discarded this idea yet.

But we need to think about how we might have just bullshitted ourselves.

One way is that this result might be dominated by a few extreme values.

Another is that this behaviour may have changed over time.

5/n

A simple way to investigate the first is to plot median returns instead of means.

If this looks substantially different, we might worry.

We see a similar pattern looking at median log returns.

6/n

Now, we might be interested in whether this effect has stayed stable over time.

It's not going to be much use to us, if all this effect was concentrated in 2019, for example.

One simple way to do this is to just split our data in subsets by date.

7/n

And when we do this, we see a consistent pattern throughout our data... which is nice.

Another useful technique is to try to squish the "shape" of those charts into a single metric we can plot every day.

That's going to give us a better idea of day to day stability.

8/n

So, let's weight each asset in inverse proportion to its distance from 20-day highs...

Our feature runs 0 to 20.

So, if an asset is:
- 0 days from its 20 day returns we'll give it a weight of 10.
- 10 days, a weight of 0
- 20 days it will get a weight of -10

9/n

And plot the cumulative sum of the the weighted next day log returns of all the assets in our data for each day.

And the extent to which it tends to go up and to the right in a smooth fashion, illustrates how consistent this effect is in our data.

10/n

We see a relatively consistent effect.

It's noisy cos its less than 3 years of daily data, it's long/short (zero net dollar exposure to crypto generally) - and, let's face it - it's pretty dumb, so you wouldn't expect more.

So we haven't been able to discard this idea.

11/n

At this point, you'd keep asking questions... do more digging.

Do we see this in individual assets by themselves? etc...

(we seem to)

Does it look distinct from the general 20d trend effect we observe?

etc etc

12/n

Then, armed with an understand of the *effect* you think is there, you might now do some simulation.

Here's a super dumb simulation (costless) of a strategy that takes equal weight long positions in every FTX spot contract that is with 5 days of its 20 day highs.

13/n

Takeaways:
- people say lots of stuff
- call bullshit on it all
- try to discard ideas as quickly as possible using simple data analysis
- if you can't discard the idea, explore the effect till you understand it
- then, and only then, simulate a trading strategy.

14/14 Fin.

For the crypto reply guy thread-skimmers coping over this...

This analysis does suggest that buying highs *IS* a good idea in crypto.

Or at least, it has been.