Deep Vector Autoregression for Macroeconomic Data

Data Science master project by Marc Agustí, Patrick Altmeyer, and Ignacio Vidal-Quadras Costa ’21

Photo by Uriel SC on Unsplash

Editor’s note: This post is part of a series showcasing Barcelona School of Economics master projects. The project is a required component of all BSE Master’s programs.

Abstract

Vector autoregression (VAR) models are a popular choice for forecasting of macroeconomic time series data. Due to their simplicity and success at modelling the monetary economic indicators VARs have become a standard tool for central bankers to construct economic forecasts. Impulse response functions can be readily retrieved and are used extensively to investigate the monetary transmission mechanism. In light of the recent advancements in computational power and the development of advanced machine learning and deep learning algorithms we propose a simple way to integrate these tools into the VAR framework.

This paper aims to contribute to the time series literature by introducing a ground-breaking methodology which we refer to as DeepVector Autoregression (Deep VAR). By fitting each equation of the VAR system with a deep neural network, the Deep VAR outperforms the VAR in terms of in-sample fit, out-of-sample fit and point forecasting accuracy. In particular, we find that the Deep VAR is able to better capture the structural economic changes during periods of uncertainty and recession.

Conclusions

To assess the modelling performance of Deep VARs compared to linear VARs we investigate a sample of monthly US economic data in the period 1959-2021. In particular, we look at variables typically analysed in the context of the monetary transmission mechanism including output, inflation, interest rates and unemployment.

Our empirical findings show a consistent and significant improvement in modelling performance associated with Deep VARs. Specifically, our proposed Deep VAR produces much lower cumulative loss measures than the VAR over the entire period and for all of the analysed time series. The improvements in modelling performance are particularly striking during subsample periods of economic downturn and uncertainty. This appears to confirm or initial hypothesis that by modelling time series through Deep VARs it is possible to capture complex, non-linear dependencies that seem to characterize periods of structural economic change.

Chart shows that improvement in performance of Deep VAR over VAR model
Credit: the authors

When it comes to the out-of-sample performance, a priori it may seem that the Deep VAR is prone to overfitting, since it is much less parsimonious that the conventional VAR. On the contrary, we find that by using default hyperparameters the Deep VAR clearly dominates the conventional VAR in terms of out-of-sample prediction and forecast errors. An exercise in hyperparameter tuning shows that its out-of-sample performance can be further improved by appropriate regularization through adequate dropout rates and appropriate choices for the width and depth of the neural. Interestingly, we also find that the Deep VAR actually benefits from very high lag order choices at which the conventional VAR is prone to overfitting.

In summary, we provide solid evidence that the introduction of deep learning into the VAR framework can be expected to lead to a significant boost in overall modelling performance. We therefore conclude that time series econometrics as an academic discipline can draw substantial benefits from further work on introducing machine learning and deep learning into its tool kit.

We also point out a number of shortcomings of our paper and proposed Deep VAR framework, which we believe can be alleviated through future research. Firstly, policy-makers are typically concerned with uncertainty quantification, inference and overall model interpretability. Future research on Deep VARs should therefore address the estimation of confidence intervals, impulse response functions as well as variance decompositions typically analysed in the context of VAR models. We point to a number of possible avenues, most notably Monte Carlo dropout and a Bayesian approach to modelling deep neural networks. Secondly, in our initial paper we benchmarked the Deep VAR only against the conventional VAR. In future work we will introduce other non-linear approaches to allow for a fairer comparison.

Code

To facilitate further research on Deep VAR, we also contribute a companion R package deepvars that can be installed from GitHub. We aim to continue working on the package as we develop our research further and want to ultimately move it onto CRAN. For any package related questions feel free to contact Patrick, who authored and maintains the package. The is also a paper specific GitHub repository that uses the deepvars package.

Connect with the authors

About the BSE Master’s Program in Data Science Methodology

Implications of Market Rating-based Segmentation on Intra-platform Competition

An application to Airbnb’s market in Barcelona. Competition and Market Regulation master project by Paul Arenas and Saúl Paredes ’21

Aerial view of Barcelona's Eixample neighborhood
Photo by Erwan Hesry on Unsplash

Editor’s note: This post is part of a series showcasing Barcelona School of Economics master projects. The project is a required component of all BSE Master’s programs.

Abstract

In recent years, large platforms have raised concerns that they may engage in anti-competitive practices that affect market competition. Therefore, analyzing the competition structure inside platforms is a relevant issue that has not been treated in much empirical research.

This study analyzes how a platform’s owner could affect the degree of competition among members of one group in the platform through biasing search results using rating classifications. In this paper, we perform an application to Airbnb‘s market in Barcelona given the particularity of rating is an unavailable searching filter to guests.

We found evidence that listing’s rating classification represents an important market segmentation in the Airbnb’s market in Barcelona that could imply a possible practice of biasing search results. Moreover, we found that the intensity of competition is differentiated by the rating-related segments, which means that these segments are concentrating competition.

Conclusions

We found an inelastic demand for Airbnb’s listings in Barcelona in a market that is divided by rating classification. In particular, our empirical results show the following two points:

First, the majority of hosts face an inelastic demand. These results are consistent under the two main models we used. From the nested logit model under rating segmentation, we found that when there is a 10% increase in price of available nights, there is an expected decrease in booked nights of 4.5%. These results imply that there is room to increase the price without reducing the revenues of the hosts.

Second, even though the rating is not available as a filter in the Airbnb web page, it creates an important market segmentation. This means that the competition between two listings that belong to the same segment is different from the competition faced by two listings that belong to different rating classifications. Moreover, we found differences in intensity of competition faced by listings that belong to different segments.

Finally, these results show that the existence of segmentation suggests that Airbnb is performing a rating-based market division. Yet the rating segmentation does not show a clear pattern of competition intensity in each group.

Connect with the authors

About the BSE Master’s Program in Competition and Market Regulation

On the Effects of Sovereign Debt Volatility: a Theoretical Model

Economics master project by Oscar Fernández, Sergio Fonseca, Gino Magnini, Riccardo Marcelli Fabiani, and Claudia Nobile

Two pairs of hands exchange euro bills
Photo by cottonbro on Pexels

Editor’s note: This post is part of a series showcasing Barcelona School of Economics master projects. The project is a required component of all BSE Master’s programs.

Abstract

We construct a theoretical Overlapping Generations (OLG) model to describe how sovereign debt crises can propagate in the economy under certain financial constraints.

In the model, households work when young and deposit their savings in exchange for a dividend, banks invest deposits in assets and government bonds. Banks, subject to legal and market requirements, invest a fixed fraction of deposits and own equity in assets. When prices of bonds fall due to perceived sovereign debt risks, banks can invest less on capital goods directly affecting the business cycle. This paper simulates the deviations from steady-state produced by a shock to government securities and provides insights into macro-prudential policy implications.

We find that a sovereign debt crisis affects young and old generations differently, with the latter facing higher fluctuations in consumption. We also find that the macro-prudential policy can be effective only at very high levels on the old, but ineffective for the younger generation.

Conclusions

This paper draws three main conclusions about the impact of a sovereign debt crisis on the business cycle within the proposed OLG theoretical framework:

  1. A decline in government bond prices leads to lower output, wages and dividend negatively affecting present and future consumption. However, this effect is different for young and old generations. In particular, the old seem to face more sudden changes and higher deviations from steady-state values when a sovereign debt crisis takes place.
  2. The proposed macro-prudential policy does not seem to offset the impact of a fall in government bonds prices on the business cycle. In fact, almost all the macroeconomic variables of interest in our theoretical model do not change significantly, relative to their steady-state values, when the supervising authority modifies the capital requirements for banks.
  3. A very aggressive policy on capital requirements (i.e. x=0.9 for the whole period) can compensate for the negative shock bonds prices have on dividends and, therefore, consumption for the old.

Connect with the authors

About the BSE Master’s Program in Economics

Personality Traits and Mental Health Outcomes: The Effect of the Covid-19 Pandemic on Young Adults in the UK

Economics of Public Policy master project by Nour Hammad, Alexandre Marin, and Ruben van den Akker ’21

Spray paint on asphalt of a smile face and the words Stay Safe
Photo by Nick Fewings on Unsplash

Editor’s note: This post is part of a series showcasing Barcelona School of Economics master projects. The project is a required component of all BSE Master’s programs.

Abstract

Mental health outcomes significantly deteriorated in the United Kingdom as a result of the Covid-19 pandemic, particularly for younger individuals. This paper uses data from the Millennium Cohort Study to investigate the heterogeneity of mental health effects of the Covid-19 pandemic on adolescents by both personality types and personality traits. Using two-step cluster analysis we find three robust personality clusters: resilient, overcontrolled, and undercontrolled.

Conclusions

  • We surprisingly find that resilient individuals, who generally have better mental health, reported larger decreases in mental health during the pandemic than both undercontrollers and overcontrollers
  • The effect seems to be driven by the neuroticism trait, such that those with higher neuroticism scores fared better than those with lower scores during the pandemic
  • Our findings highlight that personality traits are important factors in identifying stress-prone individuals during a pandemic.

Connect with the authors

About the BSE Master’s Program in Economics of Public Policy

Does Macro-Financial Information Matter for Growth at Risk Forecasting?

Finance master project by Lapo Bini and Daniel Mueck ’21

Charts show growth at risk predictions

Editor’s note: This post is part of a series showcasing Barcelona School of Economics master projects. The project is a required component of all BSE Master’s programs.

Abstract

In order to analyse whether financial conditions are relevant downside risk predictors for the 5% Growth at Risk conditional quantile, we propose a Dynamic Factor- GARCH Model, comparing it to the two most relevant approaches in the literature. We conduct an out-of sample forecasting analysis on the whole sample, as well as focusing on a period of increased European integration after the 2000s. Always, including the national financial conditions index, term structure and housing prices for 17 European countries and the United States, as down side risk predictors. We find evidence of significant predicting power of financial conditions, which, if exploited correctly, becomes more relevant in times of extraordinary financial distress. 

Conclusions

We propose a Dynamic Factor-GARCH model which computes the conditional distribution of the GDP growth rates non-parametrically, exploiting the dimensions of a panel of national financial conditions and compare it to the models of Adrian, Boyarchenko, and Giannone (2016 )and Brownlees and Souza (2021) out-of-sample.

Contrasting to our in-sample results, the out-of sample results exhibit a higher degree of heterogeneity across countries. While our model performs at least as good or better as the AR(1)-GARCH(1,1) specification of Brownlees and Souza (2021) in the long run, it produces unsatisfactory results for the one-step forecast horizon.

However, by focusing our out-of-sample analysis on a smaller sample around the period of the Great Recession, we not only outperform the other two models analysed, but also obtain strong indication of increased importance and predictive power of financial conditions.

We provide evidence that by correctly modelling financial conditions, they not only exhibit predictive ability for GDP downside risk, but also improve in-sample GaR predictions. Further, we show that they are relevant out-of-sample predictors in the long run. Finally, when focusing on periods of extraordinary financial distress, like the Great Recession, financial conditions become even more relevant. However, the right model needs to be applied in order to exploit that predictive power.

Connect with the authors

About the BSE Master’s Program in Finance

How COVID-19 affected the Catalan regional elections

Analysis by Iván Auciello Estévez ’21 and Pau Jovell Codina ’21 (Economics)

Editor’s note: this article was originaly published in the popular economics blog, Nada es Gratis (in Spanish).

In the Catalan elections held on 14 February 2021, just a few weeks after the peak of the third wave of COVID-19, voter turnout was significantly lower than in 2017 (51.3% compared to 79.1%) and the lowest in history. Despite the political context being different compared to 2017, in the run-up to the election predicted turnout remained at similar levels until the emergence of Covid-19 – whereupon it dropped sharply. In these extraordinary elections, the vote share of the parties changed which shifted the Catalan political spectrum.

We study this relationship between Covid-19 and electoral results in Catalonia. To estimate the effect of the pandemic we explore the differences in cumulative Covid-19 incidence at the municipal level and compare it with electoral outcomes controlling for economic and demographic variables (population density, percentage of over-65 years old, share of foreign population and the unemployment rate).

The most widely used model of electoral participation, that of Riker and Ordeshook (1968), considers that the decision to participate is based on a cost-benefit trade-off between the expressed benefit of voting (feeling fulfilled, considering that civic duty is fulfilled, etc.) and the cost that individuals associate with participation. This cost rose sharply, because voting implied a higher risk of contracting the disease and the inconvenience caused by anti-covid measures. Therefore, higher costs would imply lower turnout as shown by several empirical studies, which find that with small increases in cost turnout drops significantly (eg. Aldrich, 1993). We assume that the increase in cost is the same for all (although there could be differences by age, and so we control for the most vulnerable group, the over-65s), which would imply a fall in participation, but could also induce changes in the outcome. This change may be due to differences in the “sentimental” benefit of voting between voters of different parties.

Economic theory also indicates who voters choose, conditional on voting. According to “retrospective” voting theory, voters support or punish parties in government in response to their performance in a crisis such as Key (1966). In contrast, according to “prospective” voting theory, the individual votes for the party that he believes will do better or, as Leininger and Schaub (2020) argue, seeks to match the party in regional and national government for more optimal crisis management.

Effect on participation

As seen in the maps above, the areas with the highest cumulative Covid incidence also have a lower percentage of participation, especially the Barcelona metropolitan area. Our analysis shows that an increase of 100 points in cumulative incidence in the last 14 days is related to a drop of 2.6 percentage points in turnout. To understand the magnitude, this is equivalent to one extra Covid case in a municipality of 1,000 inhabitants, so we estimate that the effect of the pandemic is quite high. To understand the impact of the second and third waves of Covid-19, we have conducted the analysis with cumulative incidence measures in the last month and in the last 4 months prior to the elections. However, as the period lengthens, the effect on turnout decreases (1.3 and 0.4 percentage points respectively).

However, the political context also changed: while in 2017 the voting framework was centred on the independence process (which led to the historical record turnout); in 2021 it was the pandemic that defined the elections. However, the trigger for this change of context (and the neglect of the independence process) was the eruption of the virus. As can be seen in the graph below showing the predicted turnout for the elections to the Parliament of Catalonia. Since Covid-19 appears to be the only element of exogenous variation between municipalities when comparing the 2017 and 2021 elections, we can infer causality.

These results are in line with the theory of the myopic voter who takes into account episodes closer to the election when voting. In this case, the myopic voter is acting rationally, as they account for the actual risk at the time of the vote rather than the risk of the past few months.

Effect on results

Secondly, the elections brought a major shift in the political spectrum, as can be seen in the bar charts above. For this reason, we have not been able to include other elections, as they were not comparable with each other. We have grouped the political parties into the following ideological and identity groups:

  • Pro-independence left: ERC and CUP.
  • Pro-independence right-wing: JUNTSXCAT and PDECAT
  • Non-independence left-wing: PSC and PODEMOS
  • Non-independence right: PP, Cs, and VOX

First, we respectively analyse the pro-independence and non-independence groups, the victors (the pro-independence coalition), and the opposition. The results show that the pandemic has a positive effect on the percentage of the vote of the pro-independence parties and a negative effect on the non-independence group. This effect can be seen as a retrospective vote, showing a certain approval by pro-independence voters of how the regional government has handled the pandemic.

On the other hand, in the analysis of the groups divided into identity and ideological groups, we find that the group with the highest positive coefficient is the non-independence left – the coalition in charge of the central government at the time of the elections. This behaviour is associated with the prospective vote. With the arrival of Salvador Illa (The Spanish Minister of Health during the pandemic) as the PSC candidate, the party was reinforced and ends up as the winner of the elections. This positive effect can be seen as an attempt to align the post-pandemic recovery strategy in Catalonia with that of the rest of Spain.

In conclusion, our results suggest that COVID-19 had a significant outcome in the Catalan elections that translated into a negative relationship between the virus and turnout. In contrast, the relationship is positive between cumulative incidence and the vote of pro-independence parties and the non-independence left-wing group. At the same time, the retrospective theory seems to hold true as there has been strong support for the government in office based on their management of the pandemic.

Connect with the authors

Iván Auciello Estévez ’21 is a student in the Barcelona GSE Master’s in Economics. After graduating he plans to work as a Research Assistant at Banco de España.

Pau Jovell Codina ’21 is a student in the Barcelona GSE Master’s in Economics. After graduating he plans to work as a Research Assistant at Banco de España.

This post was edited by Ashok Manandhar ’21 (Economics).

How has public discourse on marriage equality affected change in US institutions?

Giulia Mariani ’12 (International Trade, Finance, and Development)

Gradual institutional change analyses have allowed drawing a more flexible line between stability and transformation when examining how institutions evolve over time, particularly in the absence of major critical junctures or exogenous shocks. Yet, the explanatory power of the theory has been undermined by a lack of attention to the overlapping boundaries of the modes of gradual institutional change, a relatively static model of agency, and conceptual confusion regarding what the modes of change exactly are.

In our recent article “Discursive Strategies and Sequenced Institutional Change: The Case of Marriage Equality in the United States” published in Political Studies, Tània Verge and I argue that addressing these shortcomings requires investigating the agent-based dynamics underpinning gradual institutional change and bringing to the fore the role of ideas. Indeed, ideas and discourses can have a constitutive impact in the creation, maintenance and reform of institutions, and actors strategically reframe problems and redirect solutions to influence both the process and the outcome of policy reforms.

Employing marriage equality in the United States as a case study, we show that the modes of gradual institutional change can be studied simultaneously as processes that unfold over time, often in a sequential fashion, as outcomes of these processes, and as strategies pursued by actors to steer, impede or undermine policy change.

Our results reveal that proponents and opponents of marriage equality have deployed discursive frames to legitimize institutional change to take off sequentially in a progressive direction — through the modes of “layering“ and “displacement“ — and in a regressive direction — through the mode of “conversion“.

Throughout this sequenced process, opposing actors have not only adjusted their discursive strategies to both their rivals and the targeted institutional venues, but have also shifted roles as change and status quo agents. Indeed, our study shows that the actors contesting the institutional status quo in one stage may become the actors defending it in a subsequent phase of the institutional change process, and vice versa. Thus, we argue that traditional, static conceptualizations of agency should be problematized and, rather than as resistance to gender-friendly reforms, opposition to marriage equality should be understood as a proactive mobilization to transform existing institutions.

The recent US Supreme Court decision in Fulton v. City of Philadelphia (2021) in favor of a Catholic foster care agency that refuses to work with same-sex couples, should then be understood as a victory of the years-long conservative strategy to undermine LGBT couples’ newly recognized right to marry.

Lastly, our study highlights the role of private actors as ideational entrepreneurs in the adoption and implementation of “morality policies,“ such as marriage equality. While morality policy scholars have so far predominantly examined how governmental actors shape policymaking, we show that the discursive strategies deployed by LGBT advocates, religious-conservative organizations and other private actors, such as foster care agencies, florists, and bakers, created new opportunities to influence policy debates and tip the scales to their preferred policy outcome.

Connect with the author

portrait


Giulia Mariani ’12 is a postdoctoral researcher in Political Science at Uppsala University. She is an alum of the Barcelona GSE Master’s in International Trade, Finance and Development.

What is the effect of Long-Term Care (LTC) benefits on healthcare use?

Helena Hernández-Pizarro ’12 is part of a research team that uses administrative data to estimate quality of life and health.

person holding a stress ball
Photo by Matthias Zomer on Pexels.com

This post is based on the article Ayudas a la dependencia y uso de los servicios sanitarios, ¿qué nos dicen los datos administrativos? (Nada Es Gratis, April 2021) by Helena Hernández-PizarroGuillem López CasasnovasCatia Nicodemo, and Manuel Serrano Alarcón.

Since the 2007 implementation of the “Dependency Act”, people with functional limitations in Spain can request Long-Term Care (LTC) benefits. The Act’s main objective is to improve the care and quality of life of people who have lost their autonomy. Fourteen years later, evidence on the impact of the Dependency Act in Spain on its beneficiaries remains scarce. This is partly because we need data on the quality of life of this population in order to fully evaluate its impact and we still don’t gather a suitable indicator. However, we can assess the impact of benefits by utilising data on the use of healthcare services as a proxy to estimate quality of life and health, and that is the approach we have taken in our research.

The relationship between LTC benefits and healthcare use

The effect of LTC benefits on healthcare use is not trivial and may have implications not just for the quality of life of recipients, but also for the management of healthcare services.

If access to LTC improves the health status of dependent people (for example through better treatment management, better nutrition or avoiding domestic accidents), investments in the LTC system could save healthcare providers money in the future. On the other hand, LTC benefits might increase the demand for healthcare, for example through greater health-monitoring by caregivers.

Using data on the type of healthcare service, type of admission and diagnoses, we can better understand the relationship between benefits and healthcare use, and therefore increase the efficiency in the allocation of social care and healthcare resources to design a better integrated care system.

The data

To study the effects of LTC benefits on healthcare use, we needed to gain access to data from social services and healthcare providers and then link it. As others have shown, this was not easy, but fortunately, the interest of the institutions involved in this research —CatSalut, AQuAS, the Departament de Treball, Afers socials i Famílies de la Generalitat de Catalunya (DTASF) (Labour, Family and Social services department) and CRES (UPF) — helped to facilitate this process.

Even with access to the data, measuring the effect of LTC benefits on the health system is not straightforward. Those who qualify for benefits will by definition have worse health and, regardless of the new policy, will probably make greater use of the healthcare system than those who don’t qualify for benefits. Therefore, simply comparing those applicants who receive LTC benefits with those who don’t would not help us to identify the effects of the Dependency Act.

To deal with this, we use an instrumental variable technique based on the “leniency” of the evaluators. The idea is as follows. When there is an evaluation guided by objective criteria such as when grading an exam, imposing a judicial sentence or assigning the severity of a medical case, there is always a degree of subjectivity from the person performing the assessment. It is common in research literature to consider this as a source of exogenous variation, because there is no predictable basis by which assessors should differ. This allows us to use traditional statistical methods that can identify any consequences associated with new policies. In our context, despite the fact that the assessment is based on the Dependency Assessment Scale, each examiner has a small margin of subjective interpretation. Thus, there are examiners who, on average, tend to provide slightly higher scores, so that their applicants qualify for greater LTC benefits. Since the applicant cannot choose his/her examiner, being assessed by one examiner or another affects the probability of receiving a benefit which is exogenous to the assessment process.

The results

In the two graphs below, we summarize the most important results from our research.

Figure 1 shows that access to LTC benefits decreases by 7 percentage points the probability of a group of hospitalizations considered by the medical literature as avoidable with continuous care for the elderly (such as hospitalizations for injuries, ulcers and nutritional deficiencies). This represents a 60% reduction in this type of hospitalization.

chart
Figure 1. Effect of LTC benefits on the probability of avoidable hospitalizations

Figure 2 shows that unscheduled visits to primary care decrease by almost 10 visits two years after receiving LTC benefits, a 50% reduction with respect to the mean. Our analysis by diagnosis indicates that this reduction is explained by a sharp drop in visits caused by the economic and family situation of the individual.

chart
Figure 2. Effect of LTC benefits on visits to primary care. Scheduled vs Unscheduled

Conclusions

Our results show that LTC benefits can mitigate the use of healthcare services, in line with the conclusions of previous research. Additionally, our data allows us to go one step further, identifying in detail the types of services which are the most affected.

Particularly interesting are the results relating to primary care, where LTC benefits strongly reduce visits not strictly related to health causes. It seems that reinforcing the LTC system will not only improve the quality of life of dependents and caregivers, but may also reduce the pressure on the healthcare system.

Undoubtedly, these results are important given the chronic under-financing of the LTC system, especially in a context where COVID-19 has highlighted a need for real integration between social and health care. This is just one example of how access to, and analysis of administrative data can contribute to the evaluation of public policies, facilitating better informed decision making. 

Connect with Barcelona GSE Alumni

portrait

Helena Hernández-Pizarro (ECON ’11, HEP ’12, GPEFM ’17) is a Research Fellow at the Centre for Research in Health and Economics (CRES-UPF).

Individual recourse for Black Box Models

Explained intuitively by Patrick Altmeyer (Finance ’18, Data Science ’21) through a tale of cats and dogs

Is artificial intelligence (AI) trustworthy? If, like me, you have recently been gobsmacked by the Netflix documentary Coded Bias, then you were probably quick to answer that question with a definite “no”. The show documents the efforts of a group of researchers headed by Joy Buolamwini, that aims to inform the public about the dangers of AI.

One particular place where AI has already wreaked havoc is automated decision making. While automation is intended to liberate decision making processes of human biases and judgment error, it all too often simply encodes these flaws, which at times leads to systematic discrimination of individuals. In the eyes of Cathy O’Neil, another researcher appearing on Coded Bias, this is even more problematic than discrimation through human decision makers because “You cannot appeal to [algorithms]. They do not listen. Nor do they bend.” What Cathy is referring to here is the fact that individuals who are at the mercy of automated decision making systems usually lack the necessary means to challenge the outcome that the system has determined for them. 

In my recent post on Towards Data Science,  I look at a novel algorithmic solution to this problem. The post is based primarily on a paper by Joshi et al. (2019) in which the authors develop a simple, but ingenious idea: instead of concerning ourselves with interpretability of black-box decision making systems (DMS), how about just providing individuals with actionable recourse to revise undesirable outcomes? Suppose for example that you have been rejected from your dream job, because an automated DMS has decided that you do not meet the shortlisting criteria for the position. Instead of receiving a standard rejection email, would it not be more helpful to be provided with a tailored set of actions you can take in order to be more successful on your next attempt? 

The methodology proposed by Joshi et al. (2019) and termed REVISE is an attempt to put this idea into practice. For my post I chose a more light-hearted topic than job rejections to illustrate the approach. In particular, I demonstrate how REVISE can be used to provide individual recourse to Kitty 🐱, a young cat that identifies as a dog. Based on information about her long tail and short overall height, a linear classifier has decided to label Kitty as a cat along with all the other cats that share similar attributes (Figure below). REVISE sends Kitty on the shortest possible route to being classified as a dog 🐶 . She just needs to grow a few inches and fold up her tail (Figure below).

The following summary should give you some flavour of how the algorithm works:

  1. Initialise x, that is the attributes that will be revised recursively. Kitty’s original attributes seem like a reasonable place to start.
  2. Through gradient descent recursively revise x until g(x*)=🐶. At this point the descent terminates since for these revised attributes the classifier labels Kitty as a dog.
  3. Return x*-x, that is the individual recourse for Kitty.
Animation illustrates how Kitty crosses the decision boundary
The simplified REVISE algorithm in action: how Kitty crosses the decision boundary by changing her attributes. Regularisation with respect to the distance penalty increases from top left to bottom right. Image by author.

This illustrative example is of course a bit silly and should not detract from the fact that the potential real-world use cases of the algorithm are serious and reach many domains. The work by Joshi et al. adds to a growing body of literature that aims to make AI more trustworthy and transparent. This will be decisive in applications of AI to domains like Economics, Finance and Public Policy, where decision makers and individuals rightfully insist on model interpretability and explainability. 

Further reading

The article was featured on TDS’ Editor’s Picks and has been added to their Model Interpretability column. This link takes you straight to the publication. Readers with an appetite for technical details around the implementation of stochastic gradient descent and the REVISE algorithm in R may also want to have a look at the original publication on my personal blog.

Connect with the author

portrait


Following his first Master’s at Barcelona GSE (Finance Program), Patrick Altmeyer worked as an economist for the Bank of England for two years. He is currently finishing up the Master’s in Data Science at Barcelona GSE.

Upon graduation Patrick will remain in academia to pursue a PhD in Trustworthy Artificial Intelligence at Delft University of Technology.


How can we rethink our economy for a more sustainable future?

Nils Handler ’18 presents the D\carb Future Economy Forum

I recently founded the D\carb Future Economy Forum with the goal of better informing the public debate on climate change on topics such as green growth, green macroeconomics and green innovation.

D\carb is strongly inspired by my Master’s studies at Barcelona GSE such as Antonio Ciccone’s class on Economic Growth and Albert Bravo-Biosca’s course on innovation policy.

Last week we held our virtual kick-off event, “Green Growth: Technological Innovation, Market Incentives and Investments for a Green Economy” to discuss the opportunities and risks of transitioning our economy into a sustainable future.

Our speakers were Prof. Ottmar Edenhofer, Director and Chief Economist of the Potsdam Institute for Climate Impact Research, and Prof. Cameron Hepburn, Director of the Economics of Sustainability Programme and Professor of Environmental Economics at the University of Oxford. Johanna Schiele, McCloy-fellow at the Harvard Kennedy School, moderated the event.

The event was organized jointly with the Mercator Research Institute on Global Commons and Climate Change and the Sustainability Centre of the Hertie School of Governance.

Upcoming events and more information about these topics is available on our website:

Connect with the author

Nils Handler ’18 is a PhD Student at DIW Berlin. He is an alum of the Barcelona GSE Master’s in International Trade, Finance, and Development.