DOCSEARCHbyAlgolia
Submit Docs

Wang Jinyuan's Personal Website and Portfolio

website·Indexed 29 days ago

https://website.noewang.top

A personal portfolio and academic repository of Wang Jinyuan, a student at Shanghai Jiao Tong University (SJTU) and the SJTU-Paris Elite Institute of Technology (SPEIT). The site features research on Large Language Model (LLM) fine-tuning, Machine Reading Comprehension (MRC), and automated parking strategies using Model Predictive Control (MPC). It also includes course notes on statistical physics, probability theory, and active portfolio management, as well as professional experience at the Shanghai Stock Exchange Technical Company.

Tokens
20.6K
Snippets
26
Records
153
Agent score
50%

What's inside Wang Jinyuan

  1. Overview of OpenWhisk Serverless Framework

    OpenWhisk is an open-source, event-driven, stateless FaaS (Function-as-a-Service) platform under the Apache Foundation. It runs function code and runtimes within Docker containers, utilizing the Docker engine for management, load balancing, and scaling. All components, including the API gateway, controller, and triggers, also run in Docker containers. It supports multiple languages including Nodejs, Python, Java, PHP, Ruby, Go, Rust, .NET, and Ballerina.

  2. Overview of Pre-trained Language Models for MRC

    Several Transformer-based models are used to enhance Machine Reading Comprehension (MRC) performance:

    • RoBERTa: An optimized version of BERT that removes the Next Sentence Prediction (NSP) objective and uses larger batches and learning rates.
    • ELECTRA: Uses a generator (Masked Language Model) and a discriminator to identify replaced tokens via reinforcement learning.
    • Transformer-XL / XLNet: Transformer-XL introduces a new position encoding scheme to handle long-term dependencies beyond fixed lengths; XLNet extends this using an autoregressive method to learn bidirectional context.
    • DeBERTa: Introduces 'disentangled attention' and 'mask-enhanced' training, achieving high performance on SuperGLUE with less training data than RoBERTa.

  3. Overview of Arbitrage Pricing Theory (APT)

    Arbitrage Pricing Theory (APT) is a powerful alternative to CAPM for generating expected excess returns. Unlike CAPM, which relies on a single market portfolio, APT posits that a stock's expected excess return is determined by its exposure to multiple factors. It is viewed more as an 'art' than a science because the theory does not specify what the factors are, how to calculate exposure, or how to determine the weights (factor predictions). It is particularly useful for active investment managers to generate their own expected returns based on fundamental, technical, or macro factors.

  4. Overview of Policy Fit Retrieval for Sci-Tech Enterprises

    Policy Fit Retrieval is a specialized application of Dense Passage Retrieval (DPR) used to evaluate the alignment between a sci-tech enterprise's main products/business (as described in their prospectus) and national industrial policy documents. The goal is to assist regulatory bodies in verifying the accuracy of an enterprise's technological positioning and its eligibility for listing on the STAR Market (Sci-Tech Innovation Board). Unlike simple industry classification matching, this system focuses on specific product-level alignment and requires industry-specific knowledge to handle supply chain terminology and the linguistic gap between complex corporate prose and concise policy language.

  5. Implement Structured APT Models

    Structured models assume specific relationships between variables (macroeconomic, fundamental, or market-related). There are three primary implementation approaches:

    1. Given Exposure, Estimate Factor Returns: Factor exposures are based on current stock characteristics (e.g., P/E ratio, size). The factor returns are then estimated (e.g., the BARRA model).
    2. Given Factor Returns, Estimate Exposure: Factor returns are predefined (e.g., NYSE index, Gold, Government Bond index). Stock exposures are determined by regressing the stock's excess return (relative to a benchmark like NYSE) against these factors.
    3. Hybrid Approach: A combination where exposures are first estimated using the method in approach 2, and then returns are attributed using the method in approach 1.

  6. Apply DDPG for automated parking control

    Deep Deterministic Policy Gradient (DDPG) is used to handle continuous action spaces in parking control. The architecture consists of four components: an environment, an Actor network (online and target), a Critic network (online and target), and a Replay Memory Buffer. The Actor network maps states to deterministic actions, while the Critic network estimates the Q-value. To ensure stability, the system uses a soft update algorithm for target networks and an Ornstein-Uhlenbeck (OU) process to introduce random noise for exploration.

  7. Calculate aggregate Information Ratio for multiple strategies (Additivity)

    The Fundamental Law of Active Management is additive in terms of the square of the Information Ratio. If a manager has two independent sets of predictions with different breadths and skill levels, the total IR is calculated as follows:
    IR^2 = BR_1 \cdot IC_1^2 + BR_2 \cdot IC_2^2

  8. Understand the French Higher Education structure for Engineers

    France employs a dual system of higher education: general universities and 'Grandes Écoles'. While universities primarily train researchers and teachers, the Grandes Écoles (including engineering, medical, teaching, and business schools) focus on elite professional education. Engineering schools specifically aim to produce talent useful for industrial application.

  9. Understand the relationship between philosophy and science as universality and particularity

    In Marxist philosophy, the relationship between philosophy and science is often viewed as one of universality and particularity. Philosophy studies the most general laws of motion and development of the entire world, while individual sciences focus on specific phenomena and specialized fields within that world.

  10. Architecture of a Standard Machine Reading Comprehension System

    A standard MRC system typically follows a two-module architecture inspired by dual-process theory (contextual perception and analytical cognition):

    1. Encoder: Projects natural language text into a vector space and models contextual features. It extracts multi-granularity features (word, sentence, paragraph level) and identifies salient features (named entities, syntactic structures). Modern encoders favor the Transformer architecture over RNNs (LSTM/GRU) due to parallel computing capabilities and better capture of long-range semantic dependencies.
    2. Decoder: Processes the hidden states from the encoder to project them into the task-specific output space (e.g., predicting a multiple-choice option or a start/end index for a text span).

  11. Predict Volatility using ARCH and GARCH Models

    ARCH (Autoregressive Conditional Heteroskedasticity) and GARCH (Generalized ARCH) are used to predict volatility when it exhibits predictable patterns (e.g., volatility clustering). The GARCH model assumes current volatility is influenced by: 1) a long-term mean, 2) a mean-reversion tendency, and 3) recent returns.

  12. Understand the relationship between philosophy and science as two different particularities

    An alternative view posits that philosophy and science both study specific, different parts of the world. In this framework, philosophy is not a universal umbrella but a set of specialized fields (e.g., logic, metaphysics, ethics) that coexist with natural, social, and cognitive sciences as distinct particularities.