Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Asset Price Prediction using LSTM neural network

Miguel
January 18, 2019
84

Asset Price Prediction using LSTM neural network

Beating the market is the goal of every asset manager out there, but is it possible to increase our odds using artificial intelligence?
What if applied to predicting the price of basic commodities such as wheat, rice, etc. in order to help farmers around the world to make the best decisions?
This presentation goes through some of the opportunities and challenges to do so, experimenting with a Long Short Term Memory neural network.

Miguel

January 18, 2019
Tweet

Transcript

  1. ASSET PRICE PREDICTION using LSTM neural network Javier Marti Gabriel

    Tello Albert Sanchez Joan Fontanals Hermes Valenciano
  2. Our mission Create a model that can predict the future

    potential price of an asset better than chance.
  3. Mission impossible? ⦿ Renaissance Technologies Stock Market 35 % year

    for 20-years ⦿ Google Heating / cooling 40% Energy savings
  4. Social uses / Gov uses ⦿ Oil ⦿ Water ⦿

    Power ⦿ Wheat ⦿ Soy ⦿ Coffee ⦿ Solar
  5. Other applications ⦿ Stock market prediction / Hedge funds ⦿

    Digital currencies / Bitcoin ⦿ Scientific analysis / Sea levels / Migration ⦿ Weather prediction ⦿ Inventory cost prediction ⦿ E-commerce seasonality ⦿ Engineering: energy costs and demand ⦿ Cybersecurity
  6. Limitations ⦿ Data quality ⦿ Data quantity ⦿ Choice of

    Deep Learning tools ⦿ Computational power
  7. Data we analyzed ⦿ Financial - Apple stock price ⦿

    Start_date = '2013-02-08' End_date = '2018-12-05' ⦿ Source: Yahoo Finance
  8. Pipeline and tools ⦿ LSTM neural network ⦿ Keras -

    Tensorflow ⦿ 9 layers ⦿ X epochs
  9. Lessons learned ⦿ Code may be broken ⦿ Choice of

    tools must be optimal ⦿ Uncommented code ⦿ Data quality
  10. Looking ahead >>>>> ⦿ Sentiment/News analysis ⦿ Test different algorithms

    ⦿ Window lengths ⦿ Timesteps ⦿ Change N-Epochs ⦿ Layer composition ⦿ Get more data ⦿ Ensemble - other models ⦿ Forecast different number of values ⦿ “The best predictive performance is a combination of shallow and deep learning” – Dr Francois Chollet / Keras creator