SPHS20
Project Page
Machine learning techniques offer notable benefits to financial industry, be it algorithmic trading, risk management or business automation. Due to numerical nature of finance and the accessibility to comprehensive data, deep learning approaches and their utility to the field have been widely researched. The study showcases an application of deep learning methods in finance using artificial neural networks (ANNs) and limit order books along with liquidity measures as time series data. Deep feedforward neural networks (DFNNs) are trained in a supervised learning setting to forecast mid prices, price expectations and price variances for the financial services company Garanti BBVA's stocks in 2017. Comparisons with linear models in terms of validation performance show an improvement of approximately 8 and 4 times smaller mean squared and mean absolute percentage errors, respectively, in price forecasts with DFNNs and no improvement in variance forecasts.
- Liquidity Measures:
- Definitions of liquidity measures.
- How they are obtained using limit order book data.
- Data Preparation:
- Extracting continuous session times from LOB using trade book.
- Calculating liquidity measures.
- Reducing data point precision to minute scale and forward filling.
- Normalizing data.
- Input and output of the neural network. Rolling windows and mini-batch content.
- Verifying alignment of rolling windows and their target. Examining impact of forward filling on the data.
- Model Evaluations:
- Evaluation of models trained with:
- Limit Order Book alone.
- Liquidity Measures alone.
- both Limit Order Book and Liquidity Measures.
- Evaluations of linear models used for benchmarking.
- Convergence curves.
- Results/Analysis:
- Overview of model performances.
- Histograms of errors.
- Autocorrelation analysis.
- Old Evaluations: Evaluations of models used previously.
- Changelog: Changelog.