Dynamic Time Warping を用いた高頻度取引データのLead-Lag 効果の推定Katsuya Ito
This paper investigates the Lead-Lag relationships in high-frequency data.
We propose Multinomial Dynamic Time Warping (MDTW) that deals with non-synchronous observation, vast data, and time-varying Lead-Lag.
MDTW directly estimates the Lead-Lags without lag candidates. Its computational complexity is linear with respect to the number of observation and it does not depend on the number of lag candidates.
The experiments adopting artificial data and market data illustrate the effectiveness of our method compared to the existing methods.
【DL輪読会】Efficiently Modeling Long Sequences with Structured State SpacesDeep Learning JP
This document summarizes a research paper on modeling long-range dependencies in sequence data using structured state space models and deep learning. The proposed S4 model (1) derives recurrent and convolutional representations of state space models, (2) improves long-term memory using HiPPO matrices, and (3) efficiently computes state space model convolution kernels. Experiments show S4 outperforms existing methods on various long-range dependency tasks, achieves fast and memory-efficient computation comparable to efficient Transformers, and performs competitively as a general sequence model.
Dynamic Time Warping を用いた高頻度取引データのLead-Lag 効果の推定Katsuya Ito
This paper investigates the Lead-Lag relationships in high-frequency data.
We propose Multinomial Dynamic Time Warping (MDTW) that deals with non-synchronous observation, vast data, and time-varying Lead-Lag.
MDTW directly estimates the Lead-Lags without lag candidates. Its computational complexity is linear with respect to the number of observation and it does not depend on the number of lag candidates.
The experiments adopting artificial data and market data illustrate the effectiveness of our method compared to the existing methods.
【DL輪読会】Efficiently Modeling Long Sequences with Structured State SpacesDeep Learning JP
This document summarizes a research paper on modeling long-range dependencies in sequence data using structured state space models and deep learning. The proposed S4 model (1) derives recurrent and convolutional representations of state space models, (2) improves long-term memory using HiPPO matrices, and (3) efficiently computes state space model convolution kernels. Experiments show S4 outperforms existing methods on various long-range dependency tasks, achieves fast and memory-efficient computation comparable to efficient Transformers, and performs competitively as a general sequence model.
- The document discusses methods for quantitatively measuring the brand value of B2B companies, which is an important but difficult-to-measure intangible asset. It analyzes the Eight Company Score (ECS) method, which surveys business card owners about their impressions of a company's brand.
- An empirical study found that a company's ECS has a statistically significant positive relationship with its market capitalization, indicating it contains information about intangible asset value. The relationship is stronger for B2B than B2C companies, consistent with ECS surveying business connections.
- The brand value-market cap relationship may vary by company size, industry, and whether it is primarily B2B or B2
The document summarizes a research paper on portfolio optimization using Conditional Value at Risk (CVaR). It proposes a new Regularized Multiple-CVaR (RM-CVaR) portfolio that is robust to error maximization, a drawback of traditional mean-variance optimization. The RM-CVaR approach constructs a portfolio that minimizes the maximum margin between multiple CVaR probability levels (e.g. 97%, 98%, 99%), making it less sensitive to errors in estimating return distributions than a single-CVaR portfolio. It formulates the optimization problem as a linear program to efficiently find the minimum RM-CVaR portfolio. The paper confirms through experiments that single-CVaR portfolios are
What Do Good Integrated Reports Tell Us?: An Empirical Study of Japanese Comp...Kei Nakagawa
The document analyzes integrated reports from Japanese companies using natural language processing techniques to identify differences between excellent, significantly improved, and unranked reports. Key findings include:
1) Excellent and improved reports place more emphasis on customers, employees, and long-term growth compared to unranked reports.
2) Topic modeling shows excellent and improved reports discuss customers, products, and medium-term plans while unranked reports discuss compensation more.
3) Word embedding finds excellent and improved reports consider sustainability, human resources strategies, and symbiosis while all reports sufficiently address governance.
This document provides an overview of time series prediction and cross-sectional prediction using machine learning. It discusses using supervised learning models for time series prediction to forecast future stock prices based on past price data and external variables. It also discusses using supervised learning models for cross-sectional prediction to predict relative stock returns in a universe based on criteria describing each stock. Examples of problem formulations, data types, and machine learning models for both time series and cross-sectional predictions in finance are presented.
RIC-NN: A Robust Transferable Deep Learning Framework for Cross-sectional Inv...Kei Nakagawa
The document describes a deep learning framework called RIC-NN for cross-sectional stock return prediction. It consists of three key parts:
1) A multi-factor deep learning approach to capture nonlinear relationships between stock factors and returns.
2) Weight initialization and early stopping based on rank correlation to control overfitting.
3) Deep transfer learning to augment models using knowledge from larger markets.
Experimental results on US and Pacific markets show RIC-NN with transfer learning performs best, and controlling overfitting through rank correlation outperforms epoch-based methods.
Economic Causal Chain and Predictable Stock ReturnsKei Nakagawa
The document proposes a method to predict stock returns using an economic causal chain network constructed from the text of Japanese financial statement summaries. It empirically tests this method on stocks in the TOPIX500 index from 2012-2019. The results show the method identifies lead-lag effects between stocks, with higher chain counts indicating stronger causality. Portfolios long stocks identified as effects and short stocks identified as causes outperform without using the causal network, demonstrating the method predicts short-term return reversals. The economic causal chain approach differs from prior work using supply chain networks and aims to capture higher-order causality relationships.
Stock price prediction using k* nearest neighbors and indexing dynamic time w...Kei Nakagawa
The document proposes using k*-Nearest Neighbors and Indexing Dynamic Time Warping (IDTW) to predict stock prices based on past price fluctuations. IDTW measures the similarity between stock price movements over monthly periods while accounting for price levels. k*-NN then predicts future prices based on the k nearest past patterns weighted by their IDTW distance. An empirical study found IDTW-k*NN outperformed other methods like DTW-kNN in predicting major stock indices out-of-sample, providing evidence against the efficient market hypothesis.