Detection of Net Asset Value Computation Errors in the Context of the Control of UCITS Funds' Values

The topic of the thesis is the study of NAV errors and their detection. We focus on traditional mutual funds investing in traditional markets such as equity, bonds, other types of transferable securities, and OTC derivatives which will be referred to in practice as "UCITS funds" in Europe. Since dependence and performance attribution are very central components within financial markets, the thesis covers topics such as multivariate modeling using copula functions, non-linear portfolio index benchmarking, and portfolio allocation analysis. These topics are vital components to consider for the detection of NAV errors. All these topics are placed in the context of funds' controls such as accounting, audit, or supervisory functions to strengthen investors' protection. All the studies performed in this dissertation are using data samples from Luxembourg funds.
In the first study, we obtained a unique dataset of 50 errors which enabled to set up definition and description of NAV error phenomenon. From this data, we could take out key statistical features of NAV errors by defining probabilities law replicating actual behavior of errors' material duration and maximal NAV% impact. Additionally, thanks to simulation analysis, we have been able to segregate the problem per asset classes and then provide clear research directions.
In the second study, we confirmed the presence of unit roots and heteroskedasticity within financial time series. We provided elements to associate univariate econometrics models such as ARMA-GARCH with dependence modelings such as DCC and PCC modeling, emphasizing dependence modeling with the latter.
In the third study, we built another dataset of NAV errors through the design and calibration of accounting computation of NAVs. This methodology allowed the identification of significant factors to detect a NAV error before the NAV publication. More specifically, we used logistic regressions trained on data sets of errors to provide comprehensive results for detecting error NAV in the context of equity funds. we demonstrated how different errors will behave in the case of pricing errors, subscriptions/redemptions booking errors, fee payment booking errors, and transaction booking errors. We focused on pricing and booking errors as they are the most common errors performed by asset servicers. We find our model to reach a specificity of 95% on some type of portfolio strategies for pricing/booking errors, which is impressive considering that a classification threshold of only 50% was used. We also demonstrate the significance of the factors employed for the detection, i.e., using funds returns, returns adjusted with a benchmark, and nonlinear dependence models using copula functions.