Modelling of Capital Requirements using LSTM and A-SA in CRR 3

Abstract: In response to the Great Financial Crisis of 2008, a handful of measures were taken to increase the resilience toward a similar disaster in the future. Global financial regulatory entities implemented several new directives with the intention to enhance global capital markets, leading to regulatory frameworks where financial participants (FPs) are regulated with own fund's requirements for market risks. This thesis intends to investigate two different methods presented in the framework Capital Requirements Regulation 3 (CRR 3), a framework stemming from the Basel Committee and implemented in EU legislation for determining the capital requirements for an FP. The first method, The Alternative Standardised Approach (A-SA), looks at categorical data, whereas the second method, The Alternative Internal Model Approach (A-IMA), uses the risk measure Expected Shortfall (ES) for determining the capital requirement and therefore requires the FP to estimate ES using a proprietary/internal model based on time series data. The proprietary model in this thesis uses a recurrent neural network (RNN) with several long short-term memory (LSTM) layers to predict the next day's ES using the previous 20 day's returns. The data consisted of categorical and time series data of a portfolio with the Nasdaq 100 companies as positions. This thesis concluds that A-IMA with an LSTM-network as the proprietary model, gives a lower capital requirement compared to A-SA but is less reliable in real-life applications due to its behaviour as a "black box" and is, thus, less compliant from a regulatory standpoint. The LSTM-model showed promising results for capturing the overall trend in the data, for example periods with high volatility, but underestimated the true ES.