Stock price prediction is a long-time and challenging topic in financial forecasts. Although stock markets are affected by many uncertain factors, numerous effectual approaches have been proposed to predict financial market trends using machine learning algorithms, such as Support Vector Machine (SVM) and Deep Belief Network (DBN). In this research, we propose a new approach to predict short-term stock uctuations using Soft Confidence-Weighted (SCW) Learning. The proposed method not only predicts stock trends, but also gives a quantitative measure for the stock uctuations. We consider RoC time series of a related stock class as inputs, uctuate up and down of a target stock as outputs, to train the prediction model with SCW. Some experimental results show that the approach is useful for practical purposes.

Humans have a strong tendency to belong to a group, which is called group behavior. It is said that group behavior may affect the financial market and make it inefficient. In this article, we study the relationship between group behavior and market impact by building an muti-agent based artificial market model. The results show that the maket become more inefficient with group behavior growing when exceed some threshold.

In this research, we analyzed impact on m arket efficiency of interaction between high frequency trading (HFT) and dark pool to a stock market using artificial market simulation. We introduced a market maker agent, a representative strategy of HFT, and changed its spread for order price. We also c hanged each stylized trader agents' percentage to use dark pool. The result showed that the smaller the spread of the market maker is, the more efficient the stock market becomes. We discussed the mechanism that percentage to use dark pool have a different impact to the efficiency of the market depending on the size of the ma rket maker's spread.

This article introduces a new verification method of quant model strategies as an alternative to back testing. New quant strategies will always suffer from a "chicken and egg" problem, finding initial investors who are afraid of the phenomenon that the new strategy, with good looking back test results, will sometimes fail in real investment; "betrayal of the back test". This article proposes a concept of "forward test" and "millennium test" as new tools to avoid the "betrayal of the back test" and to evaluate the forecasting capability of the new quant model. It also discusses a real application of the forward test and millennium test under the new quant strategy, displaying the strategy's capability to forecast and generate positive returns over time.

OLS (Ordinary Least Square) estimation of sample means and betas can lead to biased estimates of alpha in the presence of certain patterns of heteroscedasticity. There has been discussing on something t o do with background of modern portfolio theory for forecasting risk premium and dealing with anomaly. We demonstrate that those patterns occur in practice, and that a robust estimation process eliminates the bias with some algorithms.

The author utilized texttext-mining and deep deep-learning technic technics to forecast a monetary policy change by the BoJ BoJ. More specifically, the classifier of the BoJ's documents was developed, which picks up the document containing any trait of previouslypreviously-experienced precursor for monetary policy change. Such classifier was constructed by obtaining distributed representation of documents via Doc2Vec and feeding them into Deep Belief Network with economic timetime-series datadata. The back back-test for the period from Jan 2014 to Jan 2016 showed a fair performance of the classifier to send precursory signalsignals against two cases of additional monetary easing easing.

Vector representation of words such as word2vec is an efficient method used in text mining. However, few papers are focusing on the multilingual studies. In this paper we present the comparative study on English and Japanese resources respectively, and then we try to investigate the possible relationship between the two vector models in two languages. We first extract two word2vec models by using news resources of ten years, and then we cluster them basing on their cosine similarity for both Japanese and English respectively. Second, we extract the words related to finance and then derive two dictionaries in two languages. Finally, we make a comparison between these two dictionaries and tempt to Sentiment estimation of a cluster of one language based on similar clusters of other language.