Information Coefficient Research Articles

Application of Long Short-Term Memory (LSTM) in Multi-Factor Quantitative Investment Models

This article first introduces the relevant concepts of quantitative investment and the development history of multi-factor theory as well as the advantages of machine learning in the field of quantitative investment. After that, this article mainly establishes the quantitative stock selection model. The establishment of quantitative stock selection models mainly includes two aspects: the screening of quantitative factors and the construction of Long Short-Term Memory (LSTM) models. The screening of quantitative factors includes data preprocessing and testing the validity of single factors through the Information Coefficient (IC) method. In the establishment of the stock selection model, this article first introduces the advantages of LSTM in time series prediction, then introduces the structure of the LSTM model in detail, and then introduces the use of selected quantitative factors and the LSTM model to build the stock selection model of this article. Finally, the stock selection model constructed in this article is backtested to verify the effectiveness of the model.

CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph

Quantitative stock selection is one of the most challenging FinTech tasks due to the non-stationary dynamics and complex market dependencies. Existing studies rely on channel mixing methods, exacerbating the issue of distribution shift in financial time series. Additionally, complex model structures they build make it difficult to handle very long sequences. Furthermore, most of them are based on predefined stock relationships thus making it difficult to capture the dynamic and highly volatile stock markets. To address the above issues, in this paper, we propose Channel-Independent based Spatio-Temporal Hypergraph Pre-trained Attention Networks (CI-STHPAN), a two-stage framework for stock selection, involving Transformer and HGAT based stock time series self-supervised pre-training and stock-ranking based downstream task fine-tuning. We calculate the similarity of stock time series of different channel in dynamic intervals based on Dynamic Time Warping (DTW), and further construct channel-independent stock dynamic hypergraph based on the similarity. Experiments with NASDAQ and NYSE markets data over five years show that our framework outperforms SOTA approaches in terms of investment return ratio (IRR) and Sharpe ratio (SR). Additionally, we find that even without introducing graph information, self-supervised learning based on the vanilla Transformer Encoder also surpasses SOTA results. Notable improvements are gained on the NYSE market. It is mainly attributed to the improvement of fine-tuning approach on Information Coefficient (IC) and Information Ratio based IC (ICIR), indicating that the fine-tuning method enhances the accuracy and stability of the model prediction.

Composition Characteristics of Volatile Organic Compounds and Associated Contributions to Secondary Pollution in Shenyang Industrial Area in Summer

Based on the observation data of volatile organic compounds (VOCs) in the industrial area of Shenyang during the summer of 2019 and 2020, the composition characteristics and sources of VOCs were preliminarily studied. The ozone formation potential (OFP) and aerosol formation potential (AFP) of VOCs were also estimated using the max incremental reactivity (MIR) and aerosol formation coefficient (FAC) methods, respectively. The results showed that the average concentration of VOCs was 41.66 μg·m-3, and the proportions of alkanes, olefins, aromatics, and acetylene were 48.50%, 14.08%, 15.37%, and 22.05%, respectively. The top ten species of VOCs were primarily C2-C5 alkanes, also including acetylene, ethylene, and some aromatics, accounting for 69.25% of the total VOCs. VOCs showed obvious diurnal variation characteristics with a high concentration in the morning and evening (at 06:00 and 22:00) and a low concentration in the afternoon (11:00-16:00). According to the value of toluene/benzene (T/B) and isopentane/n-pentane, the atmosphere of the industrial area was mainly affected by vehicle exhaust emissions, solvent use, combustion sources, and LPG/NG. The total AFP of VOCs was up to 41.43×10-2 μg·m-3, and aromatics were the largest contributor. The total OFP of VOCs reached 117.59 μg·m-3, in which the alkenes contributed the most.

The statistics of time varying cross-sectional information coefficients

The information coefficient (IC), defined as the correlation coefficient between a stock return and its factor exposures predictor variables, is one of the most commonly used statistics in quantitative financial analysis. In this paper, we establish consistency and asymptotic normality of the time series average of cross-sectional sample ICs when the true underlying ICs between the risk-adjusted residual return and the standardized factor exposures are time varying. We use those results to show that the time series average of the cross-sectional sample ICs divided by its sample standard deviation converges to the ex ante expected portfolio information ratio (IR) as derived in Ding and Martin (2017). A simulation study based on a true factor model shows that the finite sample results are strikingly close to what the theory suggests. We also conduct empirical simulations using actual stock returns and quantitative factor exposures, and we find that the logarithm of the estimated IR can be explained very well by a function of the IC mean, the IC standard deviation, and the sample size in exactly the same way as predicted by our theory built on a linear factor model with time varying ICs.

Angioedema Caused by Drugs That Prevent the Degradation of Vasoactive Peptides: A Pharmacovigilance Database Study.

Angioedema results from the decreased degradation of vasoactive peptides such as substance P and bradykinin. In this study, we sought to clarify whether dipeptidyl peptidase-4 (DPP-4) and angiotensin-converting enzyme (ACE) inhibitors that suppress the degradation of substance P and bradykinin are involved in angioedema onset. We calculated information coefficients (ICs) by performing a disproportionality analysis to evaluate DPP-4/ACE inhibitor-induced angioedema using the Japanese Adverse Drug Event Report (JADER) database. No angioedema signals were detected for DPP-4 inhibitors; however, a signal was detected for ACE inhibitors (IC: 2.42, 95% confidence interval (CI): 2.19 to 2.65). Of the patients treated with DPP-4 inhibitors, four developed drug-induced angioedema in combination with ACE inhibitors, and all were taking vildagliptin. Signals were detected for enalapril (IC: 2.39, 95% CI: 2.06 to 2.71), imidapril (IC: 2.83, 95% CI: 2.38 to 3.27), lisinopril (IC: 2.28, 95% CI: 1.55 to 3.00), temocapril (IC: 1.35, 95% CI: 0.29 to 2.40), and trandolapril (IC: 1.57, 95% CI: 0.19 to 2.95). Both inhibitors inhibited the degradation of substance P and bradykinin and were thus expected to cause angioedema. However, no signal of angioedema was detected with the DPP-4 inhibitors, in contrast to some ACE inhibitors. This study found that ACE inhibitors and DPP-4 inhibitors, which inhibit the degradation of substance P and bradykinin, tended to have different effects on the onset of angioedema in clinical practice.

Association between dipeptidyl peptidase-4 inhibitor and aspiration pneumonia: disproportionality analysis using the spontaneous reporting system in Japan.

Dipeptidyl peptidase-4 inhibitor (DPP-4-Is), a kind of drug used for the treatment of diabetes, is considered to prevent the degradation of substance P that suppresses the occurrence of dysphagia. On the other hand, DPP-4 inhibitors are also known to act on the immune system. In this study, we used a spontaneous reporting system to evaluate the signals for dysphagia and aspiration pneumonia with DPP-4-Is. We calculated reporting odds ratio (ROR) and information coefficients (IC) as disproportionality analysis to evaluate DPP-4-Is induced dysphagia and aspiration pneumonia using the Japanese Adverse Drug Event Report (JADER) database. For DPP-4-Is as a class, no signals were detected for dysphagia, but the signal for aspiration pneumonia was detected at ROR 1.67 (95% confidence interval [95% CI]: 1.20 to 2.34) and IC 0.70 (95% CI: 0.21 to 1.19). For aspiration pneumonia, trelagliptin was the only drug among the DPP-4-Is for which both ROR and IC signals were detected (ROR 9.99, 95% CI: 4.10 to 24.36; IC: 1.98, 95% CI: 0.78 to 3.18). ROR signals, but not IC signals, were detected for linagliptin (ROR 2.66, 95% CI: 1.19 to 5.94; IC: 1.09, 95% CI: - 0.004 to 2.19) and sitagliptin (ROR 1.84, 95% CI: 1.04 to 3.25; IC: 0.78, 95% CI: - 0.03 to 1.58). Since DPP-4 inhibitors prevent the degradation of substance P involved in swallowing reflex, DPP-4 inhibitors were expected to prevent dysphagia and aspiration pneumonia. However, this study revealed that DPP-4 inhibitors strongly were associated with onset rather than preventing aspiration pneumonia. This result suggests that DPP-4 inhibitors may affect the immune function associated with the development of aspiration pneumonia. Furthermore, there is a possibility that the amount of DPP-4-Is used clinically cannot increase the amount of substance P in sufficient quantity to prevent aspiration pneumonia.

Information Coefficient: From a New Perspective

IC (Information Coefficient) is a widely and deeply accepted measure in active portfolio management.This paper investigates its probability properties that requires basic and in-depth research. First, this paper brings a new perspective on IC: a linear operator of a random unit vector generated by projecting returns onto a unit sphere. The new definition distinguishes IC from correlation coefficients or common statistics like t-statistics. Second, we theoretically solve the maximization of IC in expectation, and specify its almost closed-form solution with directional statistics. Third, simulation analysis reveals the influence of market condition like the stock number on IC, and emphasizes the difference between the optimal solutions and the population mean of returns. Last, empirical studies on Chinese stock market present a set of facts emerging from the unit transformed vectors of cross-sectional returns, and obtain the time series of IC in the real market with rolling windows. Our research profoundly reveals the nature of IC and sharply deepens the understanding of active portfolio management. Both the methodology and empirical study on IC suggest great potential in its further research.

Joint Distribution of Forecasts and Outcomes: Impact of Non-Normality on the Measurement of Forecasting Skill, with Applications to Analysts’ Target Prices

The purpose of this paper is to provide a detailed analysis of the joint distribution of forecasts and outcomes in the context of financial forecasting. We use Edgeworth expansions to model this joint distribution. In turn this allows us to assess the impact of non-normality, either in forecasts, or outcomes or both. This leads to multiple results; we can deduce the distribution of the forecast error; we can analyse the properties of the hit rate (H), as a statistical concept in its own right, and also in its relation to the information coefficient (IC); two tools that are used in the assessment of forecasting ability by active fund managers and financial analysts. Our paper contributes to the recent econometric literature on directional forecasting as well as the empirical literature that examines analyst performance. We find that the close link between H and IC under normality breaks down in the more general case. We provide further evidence on the richness of this approach by looking at simulation and empirical evidence.

Factor Selection with Deep Reinforcement Learning for Financial Forecasting

The selection of a relevant factor set is essential for obtaining a financial forecasting model with high information coefficient (IC). We introduce a factor selection algorithm that uses a value network to evaluate the ICs of forecasting models and a genetic algorithm to propose new factor set candidates. The value network is trained from backtesting with historical data, with the genetic algorithm proposing candidates based on the value network evaluation. Such reinforcement learning approach iteratively selects factor sets that result in higher ICs. We demonstrate a substantial performance improvement on a typical robust linear regression model. By providing an optimal input factor set, our approach can likely improve any forecasting model, including those with built-in factor selection operations. Using the validation data, we demonstrate that our algorithm is not merely fitting historical patterns. It also shows high efficacy in achieving better factor set selection for forward prediction.

Measuring intellectual capital performance of Indian banks

Purpose – The purpose of this paper is to measure the intellectual capital performance of Indian banks and established a relationship between intellectual capital and return on assets (ROA). The paper also compared the intellectual capital performance of public sector and private sector banks. Design/methodology/approach – This study is based on secondary data from the top 20 Indian banks. Ten banks were selected from each of the public and private sectors on the basis of paid-up equity capital. The analysis was made using the value added intellectual coefficient, the coefficient of variation, exponential growth rates, trend analysis, Yule’s coefficient, the coefficient of correlation, the F-test and the t-test. Findings – The study revealed that private sectors have performed relatively better regarding the creation of total information coefficient (IC). However, the ROA was still below the international benchmark of > 1 percent. The major cause of the lower IC and the reduced ROA is disproportionate to the increase in capital employed and escalating non-performing assets in the Indian banking sector. Practical implications – The study focussed on managers and identified the causes of lower performance. It proposed numerous strategies to improve the aggregate score of IC, which is closely related to bank profitability. Originality/value – This is the first study to make a comparative analysis of intellectual capital performance in public and private sector banks in India and in addition to the traditional style of measuring sectoral performance. Further, the study employed new statistical tools, such as Yule’s coefficient of association, to establish the association between performance variables.

On the Difficulty of Measuring Forecasting Skill in Financial Markets

The use of correlation between forecasts and actual returns is commonplace in the literature, often used as a measurement of investors' skill. A prominent application of this is the concept of the information coefficient (IC). Not only can the IC be used as a tool to rate analysts and fund managers but it also represents an important parameter in the asset allocation and portfolio construction process. Nevertheless, a theoretical understanding of it has typically been limited to the partial equilibrium context where the investing activities of each agent have no effect on other market participants. In this paper we show that this can be an undesirable oversimplification and we demonstrate plausible circumstances in which conventional empirical measurements of IC can be highly misleading. We suggest that improved understanding of IC in a general equilibrium setting can lead to refined portfolio decision making ex ante and more informative analysis of performance ex post. Copyright © 2015 John Wiley & Sons, Ltd.

Efficacious End User Measures—Part 1: Relative Class Size and End User Problem Domains

Biological and medical endeavors are beginning to realize the benefits of artificial intelligence and machine learning. However, classification, prediction, and diagnostic (CPD) errors can cause significant losses, even loss of life. Hence, end users are best served when they have performance information relevant to their needs, this paper’s focus. Relative class size (rCS) is commonly recognized as a confounding factor in CPD evaluation. Unfortunately, rCS-invariant measures are not easily mapped to end user conditions. We determine a cause of rCS invariance, joint probability table (JPT) normalization. JPT normalization means that more end user efficacious measures can be used without sacrificing invariance. An important revelation is that without data normalization, the Matthews correlation coefficient (MCC) and information coefficient (IC) are not relative class size invariants; this is a potential source of confusion, as we found not all reports using MCC or IC normalize their data. We derive MCC rCS-invariant expression. JPT normalization can be extended to allow JPT rCS to be set to any desired value (JPT tuning). This makes sensitivity analysis feasible, a benefit to both applied researchers and practitioners (end users). We apply our findings to two published CPD studies to illustrate how end users benefit.

The Statistics of Cross-Sectional Information Coefficient

Information coefficient (IC) is one of the most commonly used statistics in quantitative financial analysis that is usually defined as the correlation coefficient between a variable's predicted and actual values. In this paper, I derive the asymptotic distribution of the sample average cross-sectional information coefficient (IC) when the true underlying IC is time varying. I show that sample average IC divided by sample IC standard deviation approaches the ex ante expected portfolio IR as derived in Ding (2011). I extend the result to cross-sectional factor return and factor return standard deviation. Simulation result is strikingly close to what theory suggests. I also conduct empirical simulation using actual quantitative factors and the relationships between cross-sectional IC, IC standard deviation and sample size are all as predicted by a linear one factor model with time varying IC.

How Variation in Signal Quality Affects Performance

The information coefficient (IC), the correlation between forecasted and realized return, is a popular measure of signal quality. As shown in this article, variation in IC is an important source of active risk, and IC variation has an effect on optimal portfolio structure. Contrary to popular belief, the ability to take short positions in equity portfolios does not necessarily lead to superior performance. Managers who can maintain a stable IC over time will benefit from short extensions, but managers who have an unstable IC may see their performance deteriorate from increased short positions.

PSR—an efficient stock-selection tool?

In this paper we examine the simple investment factor in the Tokyo Stock Exchange (TSE). In the past, many papers have pointed out the existence of a book to market ratio of equity (PBR) anomaly in Japan. If we create portfolios with low-PBR equities, we could beat the TOPIX. Despite the asset price deflation, which started in 1991, especially in land and stock prices, the PBR was not the most efficient investment indicator in the TSE. We compare PBR, PER (earnings to market ratio) and PSR (sales to market ratio) from 1982 onwards. Over a 14-year period, the low-PSR equities beat the low-PBR and low-PER portfolio 6 times. Statistically, we calculate the IC (Information Coefficient) score between investment variables and return. Results show that the low-PSR IC score exceeds the other investment indicators, including market capitalization size. Among the 100 companies with low PSRs are industries not found on the list of low PERs and low PBRs. Thus, investors who base stock selection on PSR have a wider range of industries to choose from, which reduces the portfolio risk. We also examine the variety of selection for each factor. From the economic viewpoint, PSR is composed of asset turnover and leverage. Asset turnover is sensitive to business cycles, while leverage is significantly influenced by a company's management strategy. In economic recovery phases since the fiscal year 1970, year-to-year sales of the companies listed on the TSE first section (excluding financial and insurance sectors) went up when the business DI coincident index also increased, year-to-year.