Publications
Public Subsidies and Innovation: A Doubly Robust Machine Learning Approach Leveraging Deep Neural Networks (Link)
with Robin Sickles
Empirical Economics, 64 (6), pp 3121–3165, June 2023
Abstract: Economic growth is crucial in improving the standard of living, prosperity and welfare. In order to generate perpetual growth, factors of production must be accumulated indefinitely without diminishing returns. While factors such as physical capital (machines and labor) exhibit diminishing returns to capital, knowledge and R&D offset these effects and determine the long-run growth. Nevertheless, market imperfections drive R&D levels below the socially desired level, thus many governments intervene to increase the R&D efforts. This paper uses European firm-level data to explore the effects of public subsidies on firms’ R&D input and output. In order to deal with the complexity of the relationship between subsidies and innovation, nonparametric methods are used to learn functional relationships from the data. First, I estimate the average treatment effects based on the selection on observables assumption. Second, I consider the possible selection on unobservables, i.e. when unobserved characteristics of firms that affect innovation drive the nonrandom subsidy assignment. In this case, I use instrumental variables (IVs) to identify the local instrumental variable (LIV) curve. The identification of the LIV is obtained via double machine learning that combines a double robust IV estimation with neural networks and deep neural networks to learn functional forms from the data. I find that public subsidies increase both R&D intensity and R&D output with more pronounced effects on the R&D intensity of high technology and knowledge intensive firms. The effects of public support remain positive and significant even after accounting for treatment endogeneity.
Working Papers
Microfinance and Microfinance plus: A Doubly Robust Machine Learning Approach (Draft coming soon)
with Valentina Hartarska and Robin Sickles
Abstract: Microfinance “plus” describes Microfinance Institutions (MFIs) that offer non-financial services such as business training, health promotion, or gender empowerment in addition to the core loans and savings products. Synergies between financial and non-financial services could exist because microfinance contracts both build and use social capital and because the poor face a multitude of constraints. That is why backed up by donors, MFIs have been experimenting to establish if the poor can be better clients if also offered benefit non-financial products. However, emerging evidence from demand studies shows, at best, limited benefits from microfinance and “plus” activities. Moreover, the addition of non-core financial services like insurance (e.g., weather-based index insurance in rural areas) is unsustainable without subsidies. Yet, there is limited evidence of how the provision of microfinance plus services affects the outreach and sustainability of MFIs. To establish the ATE of microfinance “plus” we use a double machine learning method with random forests to provide a semiparametric efficient estimator of ATE. The results indicate no differences in the performance of MFIs offering core financial and microfinance plus services, possibly because much of the “plus” initiatives are subsidized. However, MFIs that offer non-core financial services together with non-financial services are serving less poor clients, suggesting a rather surprising "mission drift". Donors and policymakers encouraging MFIs to expand their menu of financial products, presumably to improve access to financial services for the poorest clients, are achieving the opposite effect - driving microfinance upmarket.
Evaluating the Impact of Regulation in Microfinance using Semi-parametric Machine Learning Methods
with Valentina Hartarska and Robin Sickles
Abstract: Microfinance institutions (MFIs) provide financial services to the poor not served by commercial banks and pursue a double bottom line – outreach and sustainability. Financial regulations are not designed with the microfinance sector in mind, and complying with regulations is not only costly but may also lead to a “mission drift,” that is, serving less poor clients. We provide evidence of the impact of regulation on the double bottom line of the microfinance industry. We use a double robust semiparametric estimation leveraging neural networks to learn the complicated functional forms of the treatment mechanism (regulation) and outcome mechanism (outreach/sustainability) from the data. Then, the trained neural networks are used to learn the treatment effects of regulation on MFIs’ performance. Results show that regulation does not affect financial results but affects the outreach of savings-and-loan MFIs. Regulation allows this group to expand its client base by 44,254 borrowers on average (across models) relative to unregulated counterparts. Simultaneously, their depth of outreach increases by about 0.13 points (relative to the average of 0.76), indicating fewer poor clients, and suggesting a mission drift. By providing separate estimates for the loan-only and savings-and-loan MFIs, we offer nuanced results to policymakers and donors who have been promoting the transformation of loan-only MFIs into commercial and regulated savings-and-loan enterprises, without understanding that this causes a mission drift.
Abstract: Predicting stock prices is not only important for finance practitioners to best allocate their assets or for academics to build better and more accurate asset pricing models, but also for the crucial implications it has about market efficiency. An efficient market would mean that all available information is incorporated in the prices and there are no arbitrage opportunities or possibilities to obtain above market returns. In this work I use recurrent neural networks and feed forward deep networks, to predict NYSE, NASDAQ and AMEX stock prices from historical data. I experiment with different architectures and compare data normalization techniques. Then, I leverage those findings to question the efficient-market hypothesis through a formal statistical test based on the method of partial derivatives suggested by White and Racine (2001) and I find evidence of an inefficient stock market. To my knowledge this is the first paper that uses neural networks combined with a particularly adapted formal statistical test to infer about market efficiency.
The Term Structure of Sovereign Credit Default Swaps and Currency Carry Trades: Two tales in Emerging and Developed Economies
with Giovanni Calice, Ming-Tsung Lin and Robin Sickles
Abstract: In this paper, we investigate the link between the term structure of sovereign credit default swaps (CDS) and the market efficiency of carry trades. Using the sovereign CDS slope as a proxy for the relative magnitude of global and country-specific sovereign risks, we document a divergent pattern of carry trade risk for developed and developing countries. In particular, we show that global uncertainty shocks in developed economies have a larger meaningful effect on carry trade crash risk, while country-level idiosyncratic shocks impact more strongly (are key determinants) on currency market risk across emerging economies. Hence, different types of sovereign risk affect individual countries carry trades. In a first step, we demonstrate that the term structure of the sovereign CDS exhibits a significant explanatory power for crash risk and for the foreign exchange risk of carry trades. In a second step, by using general factor models introduced by Kneip, Sickles, and Song (2012) (KSS), we estimate how this translates into financially important gains in terms of currency excess returns. Overall, our results are relevant to investors and to the best of our knowledge this is the first empirical evidence that CDS markets enhance currency carry trades.
Publications (Pre-PhD)
with Leze Marku, Hekuran Vrapi and Myzejen Hasani
International Journal of Green and Herbal Chemistry , 3 (2), pp 524-531, April 2014.
Abstract: The fungus Venturia inaequalis infects members of the Maloideae, and causes the disease apple scab, the most important disease of apple worldwide. With the cultivation of susceptible commercial apple cultivars, apple scab control is becoming more difficult, such that losses caused by apple scab would be about 70% if no control measures were taken. The experiment is carried out at Qerret (Puka region) Albania, during the years 2012 and chemicals tested are: - Armicarb®100 (85% KHCO3 from Helena Chemical Company, USA), 2- Kresoxim-methyl, from BASF, Belgium), 3- Thiovit jet (80% micronized sulphur, from Syngenta Agro S.A.S., France) and 4-control (non treated). The objective of this study was to evaluate the effectiveness of bicarbonates used alone or combined with horticultural oils for the control of apple scab in order to develop a successful strategy using environmentally friendly substances compatible with the organic production system.