Webinar: How to Forecast Stock Prices Using Deep Neural Networks

Join Lucena's CEO Erez Katz and learn about an innovative approach to forecasting stock prices using image representation of timeseries data.
Image recognition, hand writing recognition and speech recognition are all examples of computer vision applications used daily with uncanny accuracy in the real world. The concept behind deep learning is surprisingly easy to understand.
Through thousands of iterations of trial and error, artificial neural networks are able to classify characteristics of images and recognize new images carrying the same characteristics.
The question remains: Can the same concept apply to time-series forecasting in Finance?
Find out how alternative data in conjunction with deep learning are used to predict assets' price actions. No deep learning or quantitative experience required.

[@Lucenaresearch]

Many have asked me for reference papers to the image transformations described in the video. arxiv.org/pdf/1506.00327.pdf arxiv.org/pdf/1710.00886.pdf These two papers provide decent background and are relatively easy to follow. Enjoy! Erez

[@nut913]

This is beautiful, great information.

[@darshankachhadiya8014]

where can i find the code?

[@jackdry5607]

Thank you for making this, it's very helpful!

[@jesublade356]

Are there any article you recommend? about investigations made for this specific topic.

[@parixitbhinde]

With respect what Erez says at 26:30, given that ATR is a single value and not an interval, what does it mean to say that you expect the price to touch the upper threshold of the ATR in the next 21 days before it touches the lower threshold?

[@Lucenaresearch]

Dear parixitbhinde, The ATR stands for Average True Range which measures the volatility of a stock based on some lookback period. Assume the ATR is 2% which means based on the past month (for example) we expect the stock to oscillate between [+2%] and [-2%] relative to its current price. The label data (what we would consider match an output of 1) is trained to enable the classifier to classify the state most probable to lead a price action that the price will reach +2% first within a predetermine timeframe. The nice thing about ATR is that it takes into account market as well as idiosyncratic volatility. So today ATR may be 2% but in a week in my be 1% or 5% etc... I hope this helps.

[@parixitbhinde]

Thanks for the answer. I get it now.

[@asostrife]

Can we have more information about RP and GADF? How did you perform?

[@Lucenaresearch]

Hi AsoStrife, Absolutely -- we found the method of converting 1 dimensional timeseries to 2 dimensional rich image representation using RP to be compelling. Especially in the context of CNN and more recently Capsule Neural Networks (CapsNet). In fact, I have a follow up Webinar planned in the near future. Please stay tuned -- announcement will be forthcoming.

[@asostrife]

@@Lucenaresearch Yes off course. I mean, there is an open source code to use to convert 1 dimensional timeseries to 2 dimensional rich image?

[@Lucenaresearch]

@@asostrife Sure -- look at the pyts library - github.com/johannfaouzi/pyts/ it has GADF and other timeseries image transformations.

[@asostrife]

@@Lucenaresearch Thank you very very much.

[@Corpsecreate]

I've tried all of these techniques, but for some reason it just never works. The predictive power is horrible all the time and I dont know why.

[@Lucenaresearch]

Hi Sam - sorry to hear that you're having a difficult time. At times, it is truly a function of how predictive the data is and how you've conditioned such data for ML forecasting. I.E how is the data labeled and what you're trying to classify (objective function). What I recommend you do is, try the techniques on data you know is predictive (i.e Sine wave) and test the technique against features (time series derived from such data). Good luck!

[@urinater]

AI for stock prediction doesn’t work. If you have an iPhone or iPad, search for “Theta 45” in the App Store.

[@Corpsecreate]

@@urinater why?