Robot better than humans calling the market

[gcgang]

This Japanese Robot Calls the Market Better Than a Human - Bloomberg Business

Bloomberg article about Japanese guy w PhD in AI building a robot and correctly predicting market moves 30 days out 68% of the time.

What % would you get if you just predict "up"?

Missed the 15% down move in Jan. Doesn't seem that impressive.

Why is this a story?

 

[NW-Bound]

From the article:

The model makes a simple call -- whether the equity index will be higher or lower after 30 days -- and over almost four years it’s been right 68 percent of the time.

Don't know about the Nikkei, but the S&P 500 daily movement is up 53.6% and down 46.4% for the period of 1950-2015. So, 68% is indeed a bit better than average.

However, if the 32% that the model is wrong costs you more money than what it makes you when it is right, then it's still no good.

For example, suppose that in the 68% of the time when it is right the market only moves 1% or 2%, but in the remaining 32% when the model is wrong the market moves 10% or more in 30 days, it still costs you money.

 

[jjquantz]

But because the daily movement is biased upward, the monthly movement will be biased upward even more strongly. I wouldn't be surprised if, historically, the S&P is up about 68% of the time month over month. A quick look at year over year data shows that it's up about 80% of the time on a yearly basis, so 68% sounds about right for the historical month over month movement. I've seen this referenced somewhere, but I can't remember where.

 

[LOL!]

68% is very close to the number reported in those Dollar-Cost-Averaging versus Lump-Sum studies. LS beats DCA about 68% of the time, but not by much.

 

[photoguy]

My guess is either that just predicting "up" all the time results in a 68% correct rate for the timeframe of data used, he massively overfit to the test set, or he made some other experimental error.




Sent from my iPad using Early Retirement Forum