Assessing the relevance of an information source to trading from an adaptive-markets hypothesis perspective

Abstract

I propose a framework motivated by the Adaptive Markets Hypothesis (AMH) to analyze the relevance of a specific information source for the trading of a given security. To illustrate the applicability and advantages of this methodology, I explore the extent to which the financial statement (FS) is relevant for Credit Default Swap (CDS) trading. Specifically, I adopt a Bayesian Model Averaging approach to examine properties of the accounting metrics that enter the implied trading heuristics of the market participants. Hypothesis-testing is conducted on various horizons around the announcement dates of corporate results. The diversity of trading rules and the shift in the heuristics mix that occurred after 2008 support the AMH perspective. Overall, results show that there is a significant component of profit-motivated trading in the CDS market that relies on financial statement information, even after controlling for information transmission from alternative trading forums. Out of sample trading strategies confirm the robustness the main findings.

Keywords:

• Adaptive markets hypothesis
• Financial statement
• Bayesian model averaging
• CDS

JEL Classification:

• C44
• G14
• C11

Acknowledgement

Financial support for this project from the Research Centre of Athens University of Economics and Business is gratefully acknowledged (Project Number: EP-2662-01, EP-2264-01). I am indebted to Tassos Malliaris, Margarita Tsoutsoura, Lucia Gibilaro, Giorgio Consigli, Mike Tsionas, Dimitris Kousenidis, Seraina Anagnostopoulou, Emanuele Borgonovo and Spyros Pagratis for the valuable discussions we’ve had and their insightful suggestions. I am also grateful to the discussants and participants in the 2018 FEBS, 2017 EFMA, 17th MFS Conference and in the seminars in Bergamo University for their constructive comments. All errors are my own.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed at http://doi.org/10.1080/14697688.2020.1726438.

Notes

† Lo (2004) defines as ‘species’, competing groups of market-participants with common trading behavior. Niederhoffer (1997, Chapter 15) compares financial markets to ecosystems populated by ‘herbivores’ (dealers), ‘carnivores’ (speculators) and ‘decomposers’ (floor traders/distressed investors).

‡ These consist of the Income/Cash-flow Statements and of the Balance Sheet.

§ See the next section for the review of the literature.

¶ See for example Thaler (1993, 2005), Barberis and Thaler (2003) and Shefrin (2007).

∥ Although not frequently encountered in mainstream financial economics, approaches of this sort have long been prevalent in various sociological disciplines studying human behavior (e.g. psychology).

‡ See the excellent survey of Moral-Benito (2015) for references.

† To keep definitions consistent with market practice, I adhere to the same categorization of FS-metrics with Bloomberg. ‘Equilibrium spreads’ refer to those observed after a sufficient length of time (30 calendar days) has passed since the release of the information that we control for. The idea is that by then, any price adjustment caused by this release will have already taken place. Das et al. (2009) and Bai and Wu (2016) allow also for at least a month after the official announcement of the financial statements.

‡ Shivakumar et al. (2011) and Zhang and Zhang (2013) also find evidence of this practice, although their work concerns exclusively earnings forecasts.

† See the next section for references.

‡ In a very recent article, Manahov et al. (2019) construct artificial high-frequency traders and conduct trading simulations of 4 stocks. The authors show that natural selection among a population of heuristics performs better than conventional analytical techniques.

§ These are strategies based on signals extracted from current and past prices/returns.

¶ This section is largely based on Augustin et al. (2014) and Oehmke and Zawadowski (2015) and their excellent descriptions of the CDS market, trading motives and market participants.

∥ With respect to quant FS-based trading in particular, there has been recently an explosive growth of OR applications on topics that are intimately related to the fundamental analysis of the firm (e.g. evaluation of corporate financial performance, validation of accounting procedures, assessment of credit risk, prediction of bankruptcy). Steuer and Na (2003) and Zopounidis et al. (2015) provide in their surveys on Multi-Criteria Decision Aiding methods a fair amount of evidence corroborating this growth.

‡ Market-based models rely almost exclusively on inputs inferred from market prices (spreads, stock and option prices etc.). For the purposes of credit trading, MBMs can be distinguished into the ‘structural’ or KMV-type and the ‘reduced-form’ model categories. ‘Hybrid’ models (Duffie and Lando 2001) are affine specifications that combine both market- and accounting- based inputs.

† See for instance articles in the financial press (e.g. ‘Quant methods creep into fundamental houses’ – FT: 7/24/2011), ‘Investors switch from humans to algorithms’ (FT.com, 8/3/2016)) as well as surveys referenced in Fabozzi et al. (2008, chapter 2).

‡ Das et al. (2009) offer as an example the case of Enron: Although credit risk models calibrated on market variables predicted negligible default probability for this company, a careful analysis of its accounting data would have enabled market participants to identify that the firm’s stock price was unduly high. Given that most market-based models used this overestimated stock price as an input, it is only natural that their output was biased towards the severe underestimation of the company’s credit risk.

§ Duarte et al. (2007) discuss among other types of statistical arbitrage strategies, the so called ‘capital structure’ arbitrage that is conducted between corporate bonds and equity (stocks and options) with the help of structural models.

† Valuation ratios were not included in this analysis because they are defined with respect to the stock price. As such, they were liable to produce problems of endogeneity in some of the regressions used. In the Online Supplement of this paper I provide a short description of the Bloomberg categorization and of how these categories relate to the financial literature.

‡ In the Online Supplement I also provide a more extensive argumentation about the significance of these 3 conditions for examining market adaptability.

† See Borgonovo and Plischke 2016, for a recent survey on sensitivity analysis, methods and applications.

† See for example the discussion in Sala-I-Martin et al. (2004).

§ To calculate at a preliminary step reasonable prior model-size estimates, I use in my analysis the approach of Chalamandaris and Vlachogiannakis (2018). In that paper, the authors employed the LARS and LASSO (Least Absolute Shrinkage and Selection Operator) algorithms to select the ‘best’ model using similar panel-regression adjustments.

† The MC3 algorithm (Raftery et al. 1997) is a popular Markov Chain Monte Carlo (MCMC) algorithm that is used as a model search strategy for identifying high probability models for selection or model averaging.

‡ In effect we exclude all those models the determinant |${{\mathbf{\text{X}}}_{\mathbf{\text{r}}}}^{\mathrm{\prime }}{\mathbf{\text{X}}}_{\mathbf{\text{r}}}$| of which is less than 10⁻¹⁰ in absolute terms.

† With respect to the profitability-related metrics specifically, it is possible to argue that our evidence is supportive of the ‘insider trading’ hypothesis articulated by Acharya and Johnson (2007, 2010). It is worth noting however, that these papers examined exclusively earnings-related information whereas our framework encompasses multiple types of FS metrics.

† Financial Flexibility is almost certain to contribute at least one regressor explaining ΔCDS_[+14,+30] while the other categories contribute from zero to three regressors during the post-announcement period.

† See for example Collin-Dufresne et al. (2001), Bakshi et al. (2006), Berndt and Obreja (2010).

† See for example, Collin-Dufresne et al. (2001), Campbell et al. (2008), Tang and Yan (2010), Bonfim (2009), Berndt and Obreja (2010), Norden (2016).

‡ I used the indices of the iTraxx and CDX families (iTraxx – Europe/HiVol/Crossover, i-Traxx Asia/Japan, CDX Investment Grade/ High Volatility and High Yield).

† Profit Margin is defined as Net Income divided by Net Sales. It is therefore affected by non-recurrent items, net financial profit and taxes. Exactly for this reason it is much more difficult to analyze or predict relative to EBIT Margin.

‡ See Blanco et al. (2005), Norden and Webber (2004, 2009), Forte and Pena (2009), Zhang et al. (2009), inter alia.

§ Of course, this is not to say that the added regressors are not significant in M3: On the contrary, in results that are available upon request, both MPPs and SCPs of specific control variables (namely the lagged returns of the CDS index, stock price and implied volatility) are invariably very close to 1, confirming that they almost always enter the traders’ decision rules and they do so with the same sign.

† These results are available from the author upon request.

† ‘Suppressor’ or ‘enhancer’ variables are frequently encountered in the literature of behavioral statistics (see for example McKinnon (2008)). The inclusion of a ‘suppressor’ variable in a regression strengthens the correlation (and its significance) of the dependent variable (‘criterion’) with the independent one (‘predictor’). This usually happens because the ‘enhancer’ is capable of explaining that specific part of the predictor variable’s variance that can be considered as responsible for ‘masking’ the predictor’s real association with the criterion variable.

‡ See Mitchell and Pulvino (2012) and Bai and Collin-Dufresne (2018) inter alia.

Additional information

Funding

Financial support for this project from the Research Center of Athens University of Economics and Business is gratefully acknowledged (Project Number: EP-2662-01, EP-2264-01).