All,
I used to ask P123 to provide machine learning methods—perhaps with Python.
But Yuval and Marco have correctly pointed out that they do not restrict the data except to the extent that they are required to by their agreement with the data provider.
They even seem to welcome data scraping.
In response to feature requests, Yuval has said the data can be downloaded and whatever metric is desired can be obtained with manipulation in a spreadsheet or using Python. I am thinking of a request for the information ratio. Yuval is right about that.
I am convinced. And P123 is still working to make it easier using DataMiner, I understand. This is not a feature request of any kind.
Still, pricing data is the easiest to obtain. And in fact I used data from Yahoo! below.
And pricing data probably has no look-ahead bias.
I took pricing data from Yahoo! to obtain the excess returns for QQQ (relative to the median of a basket of ETFs) and ran a couple of different types of neural nets on the data. I divided the QQQ data in half for a validation set.
These models included the much touted Long Short-Term Memory recurrent neural net architecture.
I then ran these models (using the hyper-parameters obtained from the QQQ data) out-of-sample on a couple of other ETFs (TLT and XLE). I gave up after that and did not test this method on the entire basket of ETFs.
Bottom line. The predictions of the next month’s excess returns for these ETFs out-of-sample were useless. I will have to find another place to put my money.
I like negative results. Prevents me from wasting my time. So far this has prevented me from putting any money into machine learning ideas that use PIT time-series data.
Best,
Jim