This website uses cookies.

By clicking the "Accept" button or continuing to browse our site, you agree to first-party and session-only cookies being stored on your device to enhance site navigation and analyze site performance and traffic. For more information on our use of cookies, please see our Privacy Policy.

Estimation

Paper Session

Sunday, Jan. 8, 2017 1:00 PM – 3:00 PM

Hyatt Regency Chicago, New Orleans
Hosted By: American Economic Association
  • Chair: Brantly Callaway, Temple University

A Machine Learning Approach to Forecast Model Selection

Brian M. Scholl
,
Institute for the Study of Labor

Abstract

Model selection is often guided by exploratory analysis as well as inference-guided and history-biased user judgment; this can introduce statistical complications that are often ignored. To facilitate forecasting in limited information environments, I adapt conventions employed in machine learning to develop a new approach to forecasting economic data. This approach specifies search restrictions along with a model selection rule and statistic that conducts model selection based on out-of-sample predictive performance (BEST-RMSFE). Importantly, the approach does not rely on potentially spurious unconditional inferential statistics or the biases of forecaster judgment. I test this approach in making 1-, 2- and 3-year ahead predictions of 12 disaggregated sub-components of annual U.S. Bank Net Interest Margin data for the 1991-2010 period for a panel of 200 banks - a highly relevant context given a policy need for forecasting despite data pathologies and identification issues. In 36 separate trials, Root Mean Squared Forecast Error reductions of up to 22 percent are achieved, with reductions of over 50 percent in certain policy-relevant contexts. The proposed approach performs well vis-a-vis alternative selection approaches; other approaches occasionally select the best sequence of specifications, but BEST-RMSFE selects the best predictive sequence consistently. Extensions provide additional policy-relevant results, including scenario-conditional forecasting.

A Method for Identifying Aggregate Credit Supply and Demand Parameters Using Heteroskedascity: An Application for Brazil

Christiano Coelho
,
Central Bank of Brazil
Joao De Mello
,
Pontifical Catholic University
Marcio Garcia
,
Pontifical Catholic University
Roberto i Rigobon
,
Massachusetts Institute of Technology

Abstract

Estimation of interest-rate elasticity of aggregate credit demand and
supply is an important issue per se, as well as being crucial for calibrating
macro models, especially the frictions-augmented modern models. Such
estimation has to deal with endogeneity issues. We propose a method
based on Rigobon's (2003) identification through heteroskedasticity
approach. We implement the method using a new dataset that contains
credit flows (prices and quantities for different credit products) in Brazil at
daily frequency. Identification hinges on assumptions about the difference
in the speed of the response of demand and supply of credit, which we
justify theoretically. Results show that in Brazil consumer credit demand
is quite interest-rate inelastic, firm credit demand is elastic, and credit
supply is inelastic. These results are in line with previous anedoctal
evidence about Brazilian credit markets and suggest that the method is
successful in recovering reasonable credit demand and supply parameters.

Semiparametric Order Response Model With Correlated Thresholds: Testing Bond Over-Rating Bias

Shuyang Yang
,
Rutgers University
Zhutong Gu
,
Rutgers University
Yixiao Jiang
,
Rutgers University

Abstract

The last financial crisis calls for caveats for potential over-rating bias from credit rating agencies (CRAs), such as Moodys' because of the influence of economic interest of its shareholders. This paper focuses on estimating and testing possible rating bias by modeling CRA's bond rating decision process as a threshold-crossing ordered model with firm-specific thresholds, which can be correlated with bond and firm characteristics. To be specific, CRA compares the latent default risk index of a certain bond to a predetermined threshold for a certain rating category, which can potentially be inflated or deflated due to CRA-firm investment relationship. We propose a single-index semiparametric ordered response model with endogenous regressors to incorporate the correlated firm-specific cutoff points. With CRA-firm investment relation as a control variable and a conditional symmetry assumption on threshold structure, shift restriction condition implies identification of conditional threshold up to location and scale, which allows direct testing on the existence of bond overrating bias. The estimation proceeds in two steps: (1) semiparametric index parameters are first estimated using weighted semiparametric least-square (WSLS); (2) given estimated indices, the scaled difference of conditional thresholds are recovered from a grid search procedure. Our results confirm that there is substantial overrating bias as CRA-firm investment relation strengthens, due to the fact that actual average thresholds start to deviate from the baseline cutoff points under impartiality.

Multiple Horizon Causality in Network Analysis: Measuring Volatility Interconnections in Financial Markets

Bixi Jian
,
McGill University
Jean-Marie Dufour
,
McGill University

Abstract

Existing literature cannot provide economic and financial networks with a unified measure to estimate network spillovers for empirical studies. In this paper, we propose a novel time series econometric method to measure high-dimensional directed and weighted market network structures. Direct and spillover effects at different horizons, between nodes and between groups, are measured in a unified framework. We infer causality effects in the network through a causality measure based on flexible VAR models specified by the LASSO approach. (Non-sparse) network structures can be estimated from a sparse set of model parameters. To summarize the complex estimated network structure, I also proposed three connectedness measures that fully exploit the flexibility of our network measurement method. We apply our method to investigate the implied volatility interconnections among the S&P 100 stocks over the period of 2000 - 2015 as well as its subperiods. We find that 7 out of the 10 most influential firms in the S&P 100 belong to the financial sector. Top investment banks (Morgan Stanley, Goldman Sachs and Bank of America) have the greatest influence in the financial sector. Market connectedness is especially strong during the recent global financial crisis, and this is mainly due to the high connectedness within the financial sector and the spillovers from the financial sector to other sectors.

A Network Map of Information Percolation

Bjorn Hagstromer
,
Stockholm University
Albert J. Menkveld
,
VU Amsterdam

Abstract

We measure information percolation in securities markets for a one-security-many-markets setting. Applications range from over-the-counter dealer markets to trading in multiple electronic venues. The outcome is a network map with markets as vertices and information flows as directional edges. The approach first removes pricing errors due to, for example, liquidity trades. It then measures the information flow from A to B by the strength of B’s immediate response to A. To illustrate, we analyze information percolation in foreign exchange, both in steady state and after an event where price discovery is suddenly left to the market (Swiss franc crash).
JEL Classifications
  • C1 - Econometric and Statistical Methods and Methodology: General