Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Unsupervised machine learning with basket clusters

hackernoon.com

Post reply on HN
Print view View original

1–10 of 39 posts

Re: Unsupervised machine learning with basket clusters

#2

Post by rushabh »

So the author was able to beat S&P500 by 10% over a period from June-2016 to June-2017 using this solver. The most important question is whether the same underlying relationships will hold true for 2017-18.

This seems like a classic example of hindsight. There are many things one can tell on hindsight when the results are out. The key is if they will hold for the future too. Am I missing something?

Re: Unsupervised machine learning with basket clusters

#4

Post by justadeveloper2 »

rushabh wrote: post #2

So the author was able to beat S&P500 by 10% over a period from June-2016 to June-2017 using this solver. The most important question is whether the same underlying relationships will hold true for 2017-18. This seems like a classic example of hindsight. There are many things one can tell on hindsight when the results are out. The key is if they will hold for the future too. Am I missing something?

It doesn't matter if you can do it once. You have to do it consistently over years, which is not possible. I know about Renaissance Technologies, but I suspect they have insider information or some other angle and it's not all about their algorithms.

Re: Unsupervised machine learning with basket clusters

#6

Post by eximius »

justadeveloper2 wrote: post #4
rushabh wrote: post #2

So the author was able to beat S&P500 by 10% over a period from June-2016 to June-2017 using this solver. The most important question is whether the same underlying relationships will hold true for 2017-18. This seems like a classic example of hindsight. There are many things one can tell on hindsight when the results are out. The key is if they will hold for the future too. Am I missing something?

It doesn't matter if you can do it once. You have to do it consistently over years, which is not possible. I know about Renaissance Technologies, but I suspect they have insider information or some other angle and it's not all about their algorithms.

That is a rather large accusation. Can you restate that as something constructive or helpful to the conversation?

Re: Unsupervised machine learning with basket clusters

#9

Post by zenkat »

square90 wrote: post #8

I'm skeptical of the starmine dataset[0] being used: "The dataset is based on relationships between elements in the periodic table and public companies." This description is a bit vague. [0] http://starmine.ai/datasets/dataset_builder.html

I'm having trouble deciding if this whole post is a sly commentary on how easy it is to get machine learning wrong.

Beat the market by 10% with k-means clustering and a feature set derived from companies and chemical elements! Hahahaha no.

Re: Unsupervised machine learning with basket clusters

#10

Post by 11thEarlOfMar »

rushabh wrote: post #2

So the author was able to beat S&P500 by 10% over a period from June-2016 to June-2017 using this solver. The most important question is whether the same underlying relationships will hold true for 2017-18. This seems like a classic example of hindsight. There are many things one can tell on hindsight when the results are out. The key is if they will hold for the future too. Am I missing something?

It needs to be tested in both bull and bear markets. His time period was all bull (S&P +22%). I've been beating the S&P on a dividend play, but it's all in an up market as well, and the upward movement may be amplifying my choices on the positive side. A down market may have the opposite effect, and amplify on the negative. Also, I may have just hit a market rotation at the right time, which is not repeatable for me (since I don't know how to predict it). So I don't trust my results unless/until it also outperforms in a down market.
Post reply on HN