Using weather as a trading signal

[gusgordon]

https://www.quantopian.com/posts/using-weather-as-a-trading-signal

[ericssmith]

This comment caught my eye: "because after transaction costs it did not show enough of a gain"

I've always wondered why assumptions about transaction costs and execution speed never show up in these kinds of "studies". Anyone who has participated in markets understands they have a significant impact on results.

Also, doesn't it seem obvious that you would analyze the results based on multiple random entry & exit points rather than cherry-picked (or arbitrary) time periods?

[gfodor → ericssmith]

It's the studies that are profitable after transaction costs, hindsight bias, and cross-validation that are not published :)

[gusgordon → ericssmith]

This does take into account commission, see here: https://www.quantopian.com/help#ide-commission

I wanted to use the longest time period possible, and this was the most info I could find about forecasted cloud coverage in NYC in the past 10 years or so. It's also worth noting that this is just an example, the paper it's based off of is more in-depth. Unfortunately they concluded that transaction costs made trading based off of this strategy almost not worth it, but the main point here is that it's interesting.

[ericssmith → gusgordon]

"transaction costs made trading based off of this strategy almost not worth it"

Be interesting to know how it stacks up against other approaches in this regard.

"main point here is that it's interesting"

That it is.

[anExcitedBeast → ericssmith]

> I've always wondered why assumptions about transaction costs and execution speed never show up in these kinds of "studies". Anyone who has participated in markets understands they have a significant impact on results.

I would guess because transaction costs don't play any real part in the theory. If we only paid attention to papers that turned a profit after transaction costs, we might miss great theoretical ideas to build on.

That said, I'm neither an academic nor a competent investor, so feel free to ignore that thought.

[makeset → ericssmith]

In practice, trading costs are very steep, and a single trading signal is almost never enough to exceed them, but multiple signals can be combined to make a decision while costs stay the same. For an individual signal study like this, a no-costs assumption can be useful to see if it has any predictive value on its own.

[lotsofcows → makeset]

"but multiple signals can be combined" This sort of thing always worries me. With remarkably few disparate data sets you can get any result you want.

[pallandt]

Spurious correlation.