John Cai | Princeton University (original) (raw)

Papers by John Cai

I have implemented 2 methods for zero-shot learning on the animals with attributes dataset. Both ... more I have implemented 2 methods for zero-shot learning on the animals with attributes dataset. Both methods use SIFT features. The first method applies a 4-layer MLP while the second method uses episodic training with the Relation Network (Sung et al., 2018). Both methods outperform the 20% benchmark (required for full credit on the assignment), but with SIFT + MLP achieving higher accuracy at 25.0% while SIFT + Relation Network only achieves 23.1% accuracy. However, SIFT + Relation Network has a higher class-balanced accuracy of 24.52% vs SIFT+ RelNet's class-balanced accuracy of 24.47%.

For this assignment, I have implemented a Visual Question Answering model that has achieved an ac... more For this assignment, I have implemented a Visual Question Answering model that has achieved an accuracy of 51.97% on the validation set, which is higher than the accuracy (45%) required to receive full credit. To extract features from the questions, I adapt a pre-trained FastText model and apply a 2-layer bidirectional LSTM. To combine the features , I experiment with element-wise multiplication and a convolutional layer. Furthermore, I perform ablation studies by only feeding in the question vector and image features respectively , and using only averaged word embed-dings. I find that using averaged word embed-dings does not decrease the accuracy substantially , and also find that removing the question vector caused a larger decrease in accuracy.

In this project, we have implemented an exact pricer using a continuous analytic formula and a Pa... more In this project, we have implemented an exact pricer using a continuous analytic formula and a Partial Differential Equation (PDE) pricer for barrier options in C++ and compared the results obtained from both functions. For the PDE pricer, we have also added a setAlignment method that allows for the grid to be aligned to the barrier level. Furthermore, we have investigated the convergence properties of the PDE pricer. We find that the PDE pricer exhibits a linear order of convergence, and that the fully implicit scheme has smoother convergence (fewer fluctuations) compared to the Crank-Nicholson scheme. Moreover, we also find that the Broadie-Glasserman-Kou adjustment formula for the continuous analytic formula provides a close approximation to the discrete PDE pricer.

In this paper, we seek to improve upon state-of-the-art methods in event extraction and natural l... more In this paper, we seek to improve upon state-of-the-art methods in event extraction and natural language processing in order to predict stock price movements. We begin by extracting event tuples using OpenIE technologies. Subsequently, we use word embeddings to find general representations of the words in these event tuples. In addition, we utilize domain specific methods to identify linguistic uncertainty. We compare the results across different architectures, and also across different word/event representations. We find that the inclusion of information on uncertainty does not add much value to our predictions. We report our best predictions when we use a Convolutional Neural Network followed by a Feedforward Neural Network.

Does uncertainty help predict recessions? To control for available information, I use a dynamic f... more Does uncertainty help predict recessions? To control for available information, I use a dynamic factor model to derive common factors from a large macroeconomic panel. In addition to studying existing uncertainty measures, I introduce a novel uncertainty measure by using natural language processing to analyse news. To evaluate forecast performance, I modify tests of equal forecast accuracy for nested models. I demonstrate that many uncertainty measures help predict recessions, especially at longer forecast horizons. Furthermore, I show that my novel measure delivers a considerable increase in predicted probabilities of the 2001 recession onset over the benchmark model. 7500 words

I have implemented 2 methods for zero-shot learning on the animals with attributes dataset. Both ... more I have implemented 2 methods for zero-shot learning on the animals with attributes dataset. Both methods use SIFT features. The first method applies a 4-layer MLP while the second method uses episodic training with the Relation Network (Sung et al., 2018). Both methods outperform the 20% benchmark (required for full credit on the assignment), but with SIFT + MLP achieving higher accuracy at 25.0% while SIFT + Relation Network only achieves 23.1% accuracy. However, SIFT + Relation Network has a higher class-balanced accuracy of 24.52% vs SIFT+ RelNet's class-balanced accuracy of 24.47%.

For this assignment, I have implemented a Visual Question Answering model that has achieved an ac... more For this assignment, I have implemented a Visual Question Answering model that has achieved an accuracy of 51.97% on the validation set, which is higher than the accuracy (45%) required to receive full credit. To extract features from the questions, I adapt a pre-trained FastText model and apply a 2-layer bidirectional LSTM. To combine the features , I experiment with element-wise multiplication and a convolutional layer. Furthermore, I perform ablation studies by only feeding in the question vector and image features respectively , and using only averaged word embed-dings. I find that using averaged word embed-dings does not decrease the accuracy substantially , and also find that removing the question vector caused a larger decrease in accuracy.

In this project, we have implemented an exact pricer using a continuous analytic formula and a Pa... more In this project, we have implemented an exact pricer using a continuous analytic formula and a Partial Differential Equation (PDE) pricer for barrier options in C++ and compared the results obtained from both functions. For the PDE pricer, we have also added a setAlignment method that allows for the grid to be aligned to the barrier level. Furthermore, we have investigated the convergence properties of the PDE pricer. We find that the PDE pricer exhibits a linear order of convergence, and that the fully implicit scheme has smoother convergence (fewer fluctuations) compared to the Crank-Nicholson scheme. Moreover, we also find that the Broadie-Glasserman-Kou adjustment formula for the continuous analytic formula provides a close approximation to the discrete PDE pricer.

In this paper, we seek to improve upon state-of-the-art methods in event extraction and natural l... more In this paper, we seek to improve upon state-of-the-art methods in event extraction and natural language processing in order to predict stock price movements. We begin by extracting event tuples using OpenIE technologies. Subsequently, we use word embeddings to find general representations of the words in these event tuples. In addition, we utilize domain specific methods to identify linguistic uncertainty. We compare the results across different architectures, and also across different word/event representations. We find that the inclusion of information on uncertainty does not add much value to our predictions. We report our best predictions when we use a Convolutional Neural Network followed by a Feedforward Neural Network.

Does uncertainty help predict recessions? To control for available information, I use a dynamic f... more Does uncertainty help predict recessions? To control for available information, I use a dynamic factor model to derive common factors from a large macroeconomic panel. In addition to studying existing uncertainty measures, I introduce a novel uncertainty measure by using natural language processing to analyse news. To evaluate forecast performance, I modify tests of equal forecast accuracy for nested models. I demonstrate that many uncertainty measures help predict recessions, especially at longer forecast horizons. Furthermore, I show that my novel measure delivers a considerable increase in predicted probabilities of the 2001 recession onset over the benchmark model. 7500 words