Finance studies often employ heterogeneous datasets from different sources with different structures and frequencies. Some data are noisy, sparse, and unbalanced with missing values; some are unstructured, containing text or networks. Traditional techniques often struggle to combine and effectively extract information from these datasets. This work explores representation learning as a proven machine learning technique in learning informative embedding from complex, noisy, and dynamic financial data. This dissertation proposes novel factorization algorithms and network modeling techniques to learn the local and global representation of data in two specific financial applications: analysts’ earnings forecasts and asset pricing.
Financial analysts’ earnings forecast is one of the most critical inputs for security valuation and investment decisions. However, it is challenging to fully utilize this type of data due to the missing values. This work proposes one matrix-based algorithm, “Coupled Matrix Factorization,” and one tensor-based algorithm, “Nonlinear Tensor Coupling and Completion Framework,” to impute missing values in analysts’ earnings forecasts and then use the imputed data to predict firms’ future earnings. Experimental analysis shows that missing value imputation and representation learning by coupled matrix/tensor factorization from the observed entries improve the accuracy of firm earnings prediction. The results confirm that representing financial time-series in their natural third-order tensor form improves the latent representation of the data. It learns high-quality embedding by overcoming information loss of flattening data in spatial or temporal dimensions.
Traditional asset pricing models focus on linear relationships among asset pricing factors and often ignore nonlinear interaction among firms and factors. This dissertation formulates novel methods to identify nonlinear asset pricing factors and develops asset pricing models that capture global and local properties of data. First, this work proposes an artificial neural network “auto enco der” based model to capture the latent asset pricing factors from the global representation of an equity index. It also shows that autoencoder effectively identifies communal and non-communal assets in an index to facilitate portfolio optimization. Second, the global representation is augmented by propagating information from local communities, where the network determines the strength of this information propagation. Based on the Laplacian spectrum of the equity market network, a network factor “Z-score” is proposed to facilitate pertinent information propagation and capture dynamic changes in network structures. Finally, a “Dynamic Graph Learning Framework for Asset Pricing” is proposed to combine both global and local representations of data into one end-to-end asset pricing model. Using graph attention mechanism and information diffusion function, the proposed model learns new connections for implicit networks and refines connections of explicit networks. Experimental analysis shows that the proposed model incorporates information from negative and positive connections, captures the network evolution of the equity market over time, and outperforms other state-of-the-art asset pricing and predictive machine learning models in stock return prediction.
In a broader context, this is a pioneering work in FinTech, particularly in understanding complex financial market structures and developing explainable artificial intelligence models for finance applications. This work effectively demonstrates the application of machine learning to model financial networks, capture nonlinear interactions on data, and provide investors with powerful data-driven techniques for informed decision-making.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
