Project Meeting (20/10/2017)

Topics discussed

Literature Review Basic Structure (for now)

• Macroeconomics
Importance of macroeconomic forecasting. Analysis of the data set and different parameters.

• Time Series
Analysis of standard forecasting approaches. Discussion on seasonality, stationarity and other characteristics of time series data.

• Deep Learning
Discussion of the emerging field of deep learning - advantages and disadvantages. Analysis of deep learning approaches.

• Macroeconomic Forecasting using Deep Learning
Connecting the above topics and analysis on the work done in the field so far.

Missing Data

Dealing with missing values of parameters in the data should be done with caution. Possible solutions include leaving out data points with missing values or imputation techniques to estimate missing values. Leaving out data points may cause disruption in the continuity ofe the time series.

Reading Resources

• Using R for Time Series Analysis - Classical methods for time series analysis and how to use them in R.
• Parameter Optimization of MultilayerPerceptron Network for Stock FinancialForecastingGianluca Cremi - Past project on financial forecasting using neural networks. Can use the models described in the project as a starting point for the neural network architecture. 

Neural Network Inputs and Outputs

Potential inputs for the neural network will consist of four values for each parameter in order to overcome effects of seasonality in the data. If using CNN, experiments have to be performed in order to find out which parameters should be "close" to each other (be part of the same convolution at some point). If using RNN might not need to input four  parameter values at a time since the recurrent connections will enable the network to take into account past values.

Potential activation functions for the output layer are linear and tangent since they are both defined in the range (-∞; +∞).

Tasks Completed

• Read more on deep learning - joined deep learning reading group and revised basic topics such as Gradient Descent algorithms, MLP, Activation Functions.
• Found several research papers on financial forecasting, time series prediction and deep learning and read some of them. Should read the rest.
• Created first neural network in Keras and achieved good results on a classical data set.
• Explored data exploration and data preprocessing libraries in python - Seaborn, scikit-learn, matplotlib.

Tasks for Next Week

• Read deep learning book what's relevant from part 1 and part 2 at least chapters 6-10.
• Read Using R for Time Series Analysis 
• Read Parameter Optimization of MultilayerPerceptron Network for Stock FinancialForecastingGianluca Cremi and implement NN in Keras.
• Read more papers.
• Read any additional materials for deep learning reading group.
• Write up literature review outline in Latex and further break it down into smaller topics.
• Figure out correct formatting for project report (fonts, margins etc.) and modify document accordingly.