ACSS-PSL Institute Research Seminar

Interpretable Machine Learning for genomic, metagenomic and other Omics data

Prison, Mental Health and Family Spillovers

Coauthors: M. Bhuller and K. Løken

Abstract:

Does prison cause mental health problems among inmates and their family members? Correlational evidence shows that the prevalence of mental health problems is much higher in the inmate population than in the general population, but it remains silent on causality. We exploit the strengths of the Norwegian setting and the richness of the data to accurately measure the impacts of incarceration on the health of the defendants and their family members. First, we use an event-study design around the case decision event. The event study is complemented with an instrumental variable (IV) strategy that takes advantage of the random assignment of criminal cases to judges who differ in their stringency. Both methods consistently show that the positive correlation is misleading: incarceration lowers the prevalence of mental health disorders among defendants as measured by mental-health related visits to health care professionals. We further demonstrate that this effect lasts long after release and is unlikely to be driven by a shift in health care demand. Family members also experience positive spillovers on their mental health, especially spouses.

When are Google data useful to nowcast GDP: An approach via preselection and shrinkage

Coauthor: Anna Simoni; CNRS - CREST

Abstract:

Alternative data sets are widely used for macroeconomic nowcasting together with machine learning-based tools. The latter are often applied without a complete picture of their theoretical nowcasting properties. Against this background, a theoretically grounded nowcasting methodology is proposed that allows researchers to incorporate alternative Google Search Data (GSD) among the predictors and that combines targeted preselection, Ridge regularization, and Generalized Cross Validation. Breaking with most existing literature, which focuses on asymptotic in-sample theoretical properties, we establish the theoretical out-of-sample properties of our methodology and support them with Monte-Carlo simulations. We apply our methodology to GSD to nowcast GDP growth rate of several countries during various economic periods. Our empirical findings support the idea that GSD tends to increase nowcasting accuracy, even after controlling for official variables, but that the gain differs between periods of recessions and of macroeconomic stability.

Explainable Performance

Coauthors: Sullivan Hué, Aix-Marseille University - Aix-Marseille School of Economics, Christophe Hurlin, University of Orléans, Christophe Pérignon, HEC Paris

Abstract:

We introduce the XPER (eXplainable PERformance) methodology to measure the specific contribution of the input features to the predictive or economic performance of a model. Our methodology offers several advantages. First, it is both model-agnostic and performance metric-agnostic. Second, XPER is theoretically founded as it is based on Shapley values. Third, the interpretation of the benchmark, which is inherent in any Shapley value decomposition, is meaningful in our context. Fourth, XPER is not plagued by model specification error, as it does not require re-estimating the model. Fifth, it can be implemented either at the model level or at the individual level. In an application based on auto loans, we find that performance can be explained by a surprisingly small number of features, XPER decompositions are rather stable across metrics, yet some feature contributions switch sign across metrics. Our analysis also shows that explaining model forecasts and model performance are two distinct tasks.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4280563

The Covid-19 Crisis: an NLP Exploration of the French Twitter Feed (February-May 2020)

Abstract:

The Covid-19 pandemic offers a spectacular case of disaster management. In this literature, the paradigm of participation is fundamental: the mitigation of the impact of the disaster, the quality of the preparation and the resilience of the society, which facilitate the reconstruction, depend on the participation of the populations. Being able to observe and measure the state of mental health of the population (anxiety, confidence, expectations, ...) and to identify the points of controversy and the content of the discourse, are necessary to support measures designed to encourage this participation. Social media, and in particular Twitter, offer valuable resources for researching this discourse. The objective of this empirical study is to reconstruct a micro history of users’ reactions to the pandemic as they share them on social networks. The general method used comes from new processing techniques derived from Natural Language Processing (NLP). Three analysis methods are used to process the corpus: analysis of the temporal evolution of term occurrences; creation of dynamic semantic maps to identify co-occurrences; analysis of topics using the SVM method. The main empirical result is that the mask emerges as a central figure of discourse, at least in the discourse produced by certain social media. The retrospective analysis of the phenomenon allows us to explain what made the mask a focal point not only in conversation, but also in behaviors. Its value resides less in its functional qualities than in its ability to fix attention and organize living conditions under the threat of pandemic.

Measuring Crime Reporting and Incidence: Method and Application to #MeToo

Abstract:

This paper studies the Me Too movement’s effects on sex criminality. As many victims do not report to the police, a long-standing empirical challenge with reported crime statistics is that they reflect variations in victim reporting and crime incidence. To separate both margins, I develop a duration model that studies the delay between the incident’s occurrence and its report to the police. The model accounts for unobserved heterogeneity, never-reporters, and double-truncation in the data. I apply it to the police records of large US cities. Contrary to the widespread view that #MeToo was a watershed moment, I find that sex crime reporting had already been increasing for years before its sudden mediatization in October 2017. Nonetheless, the movement had a positive, persistent impact on victim reporting, particularly for juveniles, racial minorities, and victims of misdemeanors and old crime incidents. The increase in reporting translates into drastically higher probabilities of arrest for sex offenders. Using reported non-sexual crimes as a control group, difference-in-differences estimates suggest the movement also had a sizable deterrent effect.

Bet on a bubble asset ? An optimal portfolio allocation strategy

Coauthors: Gilles de Truchis, Elena Dumitrescu, Sebastien Fries

Abstract:

We discuss portfolio allocation when one asset exhibits phases of locally explosive behavior. We model the conditional distribution of such an asset through mixed causal-non-causal models which mimic well the speculative bubble behaviour. Relying on a Taylor-series-expansion of a CRRA utility function approach, the optimal portfolio(s) is(are) located on the mean-variance-skewness-kurtosis efficient surface. We analytically derive these four conditional moments and show in a Monte-Carlo simulations exercise that incorporating them into a two-assets portfolio optimization problem leads to substantial improvement in the asset allocation strategy. All performance evaluation metrics support the higher out-of-sample performance of our investment strategies over standard benchmarks such as the mean-variance and equally-weighted portfolio. An empirical illustration using the Brent oil price as the speculative asset confirms these findings.

The Origins of Commodity Price Fluctuations

Coauthors: Sarah Mouabbi (Banque de France) et Adrien Rousset Planat (London Business School)

Abstract:

We build novel indexes of commodity price developments by simulating news-reading. Our proposed computer-based, narrative approach is flexible, unified and spans the global commodity market, including energy, industrial and precious metals, and agricultural commodities. Empirical evidence and human readings of news articles indicate that our indexes capture commodity-price supply and demand components. Index-peaks track the post-crisis collapse of commodity markets, other market-specific developments, as well as the recent COVID-19 crisis. The richness of news content allows us to decompose the supply and demand indexes into a number of key determinants that shaped commodity markets since the beginning of the 21st century, including business cycle effects, geopolitical risk, natural disasters, and climate change considerations. Preliminary results reveal that the nature of commodity price movements matters for macroeconomic outcomes, firms’ decisions, and asset prices.