Quantitative Trading Wed, 04 Dec 2019 19:41:00 GMT language
By Akshay Nautiyal, Quantinsti
The tails of an empirical return distribution are always thick, indicating lucky gains and enormous losses are more probable than a Gaussian distribution would suggest.
Empirical distributions of assets show sharp peaks which traditional models are often not able to gauge.
A few tweaks which reduced errors for both Generator and Discriminator were 1) using a different learning rate for both the neural networks. Informally, the discriminator learning rate should be one order higher than the one for the generator. 2) Instead of using fixed labels like 1 or a 0 (where 1 means “real data” and 0 means “fake data”) for training the discriminator it helps to subtract a small noise from the label 1 and add a similar small noise to label 0. This has the effect of changing from classification to a regression model, using mean square error loss instead of binary cross-entropy as the objective function. Nonetheless, these tweaks have not eliminated completely the suboptimality and mode collapse problems associated with recurrent networks.
2) Nicolas Ferguson has translated the Kalman Filter codes in my book Algorithmic Trading to KDB+/Q. It is available on Github. He is available for programming/consulting work.
3) Brain Stanley at QuantRocket.com wrote a blog post on "Is Pairs Trading Still Viable?"
4) Ramon Martin started a new blog with a piece on "DeepTrading with Tensorflow IV".
5) Joe Marwood added my book to his top 100 trading books list.
6) Agustin Lebron's new book The Laws of Trading contains a good interview question on adverse selection (via Bayesian reasoning).
7) Linda Raschke's new autobiography Trading Sardines is hilarious!