Forecasting Macroeconomic Parameters with Deep Learning Neural Networks

Project Meeting (24/11/2017) Topics Discussed Data augmentation Data augmentation could be useful if the amount of data is not enough. Different techniques of the data augmentation discussed were: Interpolation Using equations to recursively predict one variable while performing one step ahead predictions for more. Adding random noise to existing data At this stage it is too early to determine weather or not data augmentation is required and could be useful. It remains a possibility for future experiments. Also transfer learning could be performed, training the network first on the augmented data and then retraining on the observed data. Transfer Learning Transfer learning could be applied using augmented/simulated data, data from other countries and data from EA countries for pretraining. The most promising option of the above would be using EA individual country data to pretrain and then train on the EA data. It is still early to decide we...

Machine Learning Strategy Notes Notes on the coursera course Structuring Machiene Learning Projects I recently started. Intro Orthogonalisation - varying one hyper parameter affect exactly one metric. Examples of orthogonal hyperparameters: Metric Hyperparameter Fit on training set network size, optimisation algorithm Fit on validation set regularisation, bigger training set Fit on test set bigger validation set Early stopping is not very orthogonal because it affect both training and validation fit. Defining a goal Evaluation metric It is a good practice to have a single number evaluation metric because it makes it easier to compare different models. This might mean combining several evaluation metrics using an average/harmonic mean or other approach. Satisficing evaluation metric Evaluation metric which needs to be only below within a certain interval. As long as the metric is within the specified interval there is no...

Project Meeting (17/11/2017) Topics Discussed Report Structure Intro Lit Review Financial time series - description/properties preprocessing modelling - arima models AI models Lagramge Mathew Butler ’s paper Neural Networks for modelling time series (this should be half of the lit review) general NN … Problem Analysis Research Hypothesis - mention using arima as baseline Software Engineering Requirements Gantt chart Ethics Statement? Design and implementation - ARIMA experiments Results and Evaluation Put results from ARIMA experiments here Appendix - all results go here Writing workflow Writing things up should ideally happen after immediately after performing experiment. Experiments generally should not be repeated until project is pretty much done. while report not finished Perform Experiment Write Results and Evaluation Write Design and Implementation Write Problem Analysis Write Literature Review Tasks Comp...

Topics Discussed Experiments performed Discussed the experiment results and how to interpret them. Even if overall forecast is not very accurate, determining the direction in which the variable will go is still valuable. Should produce a plot (3D / heatmap) of hyper-parameter grid search. Results from this paper Useful to see “recursive” prediction results to compare with the ones obtained with neural networks. A prediction of two years ahead is very good. Maybe possible to motivate use of deep learning for forecasting based on results from non-linear equation discovery models. Further experiments Should create an ARIMA and use it as a benchmark to compare results from neural nets. Further univariate variable experiments should be performed using seasonal and trend decomposition on input data and first order difference. Tasks Competed [x] Perform experiments with simple NN. [x] Establish experimentation workflow. Tasks for Next Week [] Create b...

Attached to this post are the results from initial experiments with the EA data set. The csv files contain details of all runs made with the specific variable. The png files are plots of the prediction of the best model found for the particular input. The numbers in file names indicate how many lags were used for the experiment(e.g. 2 means inputs were and and output was )

Topics Discussed Software Engineering Aspect of Project Need to have software engineering aspect of project. Potential ideas to be explored: Requirements engineering Unit Testing Planning - Gant Chart Report Structure Potential page limits for individual sections were discussed: Introduction: 2 Lit Review: 10 Problem Analysis: 5 Design and Implementation: 10 Results and Evaluation: 10 Conclusion: 3 Neural Network Initial experiments with a simple architecture need to be conducted. Experimentation workflow should be established in order to enable fast reliable prototyping. Unit tests can be used to ensure reliability of experiments. Statistical tests will be used to compare the performance of different architectures (ANOVA).  Tasks Completed Learn about stationarity, trends and seasonality. Analyse data - perform simple regression, figure out if it's stationary or not, identify trends and seasonality. Write up analysis. Read  Using R for Tim...

Data Analysis Exploratory data analysis was performed as agreed upon in last weeks meeting. The report can be seen below.