Quant Competition 2023

Summary of the contest

Second Edition, academic contest open to multiple Master programs in Romania (1st and 2nd year students), in the area of Data Science / Machine Learning / Econometrics, including but not limited to:

• ASE Cibernetica – all masters
• Universitatea Bucuresti – Master Data Science
• Master Artificial Intelligence
• Master Probabilitati si Statistica in Finante
• Univ Babes-Bolyai Cluj – Master Econometrie si Statistica Aplicata
• Univ Babes-Bolyai Cluj – Master Data Science for Industry and Society
• Univ Babes-Bolyai Cluj – Master Analiza Datelor Complexe
• Univ A.I. Cuza Iasi – Master Data Mining (Cibernetica si Statistica)
• Universitatea de Vest Timisoara – Master Big Data (data science, analytics and technologies)

Students are given a real-world professional assignment and will be assessed jointly by their Professor and by a jury of Deloitte & BT professionals.

The outcome of the program for students include financial awards for top 3 participants (below the net amounts), as well as multiple invitations for collaboration (hiring) in the Risk Advisory team at Deloitte and at Banca Transilvania:

• 1st place: €2000
• 2nd place: €1500
• 3rd place: €1000

Assignment - Overview and Dependent variable calculation

• Context: Time series models are often used in the estimation of forward-looking component of probability of default models used in quantification of losses according to the IFRS 9 standard. Their role is very important in capturing the effect of macroeconomic factors and other exogenous risk drivers that may affect the level of credit risk of a loan portfolio. In order to build such a model a bank usually calculates the observed PD or the so-called observed default rate (DR). The outcome of the DR is 12 months, this is equivalent to a PD for the next 12 months.

• The assignment of this contest will challenge you to make a model that is suitable to quantify the forward-looking component, you will start with the calculation of the target/dependent variable and potentially end with producing the output of such a model, the forecasts.

• Database: information at client level. It will be provided prior to the start date, via a Deloitte Shared Server. This will be the source of calculation for the dependent risk driver.

• Macroeconomic info: The independent risk drivers should be of macroeconomic nature, it is public information. The participants are expected to gather these risk drivers from the internet (ex: Eurostat, INSSE, BNR, World Bank etc.)

• Main assignment (“the semi-finals”):

The model can be developed in SAS/R/Python/Eviews/SPSS/Excel/etc. The exercise will take place in teams of 2 students and has two parts:

1. Calculate the dependent variable to be modeled from the Database: observed default rate with outcome of 12 months.
   • At a given reporting date the observed default rate is calculated as: number of defaulted clients in the next 12 months divided by the total number of performing clients at reporting date. For example, if there are 150 performing (non-default) clients in December 2021 (e.g. a reporting date) out of which 30 have at least one default event during Jan-Dec2022, then the observed default rate at December 2021 is 30/150=20%.
2. Develop a time-series statistical model for the estimation of the default rate using macroeconomic risk drivers. The dependent variable is the observed default rate from point 1). You can identify and download drivers from INSSE, BNR, Eurostat, etc.
   • Potential models: linear regression, autoregressive (AR), autoregressive moving average (ARMA), autoregressive with distributed lags (ARDL), models to address seasonality endogenously (e.g. SARIMAX), random forest, neural networks, RNN, LSTM, etc.
   • Assess model performance through specific validation tests: p-values, R-squared, relevant business sign of macro drivers (e.g. negative GDP sign), best linear unbiased estimator BLUE tests (no collinearity, stationarity of drivers, homoscedasticity, no autocorrelation, normality of errors, linear functional form). If BLUE tests fail, we expect an explanation for the failure, it is not necessary that all BLUE tests are passed. For non-linear models, apply BLUE tests where appropriate and discuss, and also apply relevant techniques for model explainability.
   • Bonus points: (1) generate a second “alternative” model using a different modeling technique (e.g. if the main model is a linear regression, an alternative model can be AR), (2) model with more than 2 relevant macroeconomic drivers, (3) obtaining at least a 70% R-squared or equivalent with all validation tests accounted for on linear regression (4) usage of transformed risk drivers (e.g. LAGs, differences, logs, etc. and determining their optimal combination in the model); (5) conduct scenario analysis to study the effect of shocks on predicted default rates, determine an appropriate size of the shock and provide economic interpretations;
   • Prepare model documentation (minimum 5 slides in PowerPoint).

• NOTE: You do not need to apply all modelling techniques or use all types of drivers. You are free to choose one or more from the requirements above. Apply the relevant statistical tests, show your work and ensure it is reproducible.

Additional assignment (“the finals”): will be communicated to qualified teams. The work in the finals will be done individually by each student.