Detecting Financial Bubbles: Dynamical and Fundamental Approaches

Financial bubbles are notable for disruptive events and severe financial consequences that adversely affect economic and financial activities. Consequently, much research and experimentation attempt to understand, identify, and forecast potential bubbles, and to mitigate related financial and economic risks. However, despite decades of inquiry and analysis, researchers still do not understand well the formation and termination of financial bubbles. In fact, they cannot even agree on a common definition, or even whether they are just patterns retrospectively classified with huge hindsight biases, such as being recognized only after they burst. This research focuses on how to detect price-related and fundamental-related financial bubbles in real time, and in particular before they end. Chapter one provides theoretical backgrounds and a social bubble framework, and describes the log-periodic power-law singularity (LPPLS) model. Chapter two focuses on macro-level bubbles, while Chapters three and four concentrate on financial bubbles from price and fundamental perspectives, using the LPPLS model and machine learning approaches. Conclusions are outlined in Chapter five. Chapter one of this thesis first presents a general introduction to market efficiencies and inefficiencies such as the Efficient Market Hypothesis and its critics, and introduces some alternative hypotheses. It then proposes a general Social Bubble framework, that incorporates financial bubbles. Based on economic scale, we further separate financial bubbles into two groups: macro-level bubbles (economic booms and busts), and micro- level bubbles, which contain three subclasses: price, valuation, and fundamental. Chapter two is based on a working paper, analyzing 20 financial bubbles in global history in support of the proposed ‘Bubble Triangle Theory’. It is posited that all 20 bubbles share three essential elements: (1) Disruptive Novelty (new product, new market, change of economic policy, or catastrophic event), (2) Abundant Liquidity and Credit (domestic credit expansion and international capital flows); and (3) the ‘Social Bubble’ spirit. Chapter three uses the LPPLS methodology to diagnose price bubbles. This chapter consists in two papers. The first paper, released in English and German, focuses on the market index level of price bubbles. It uses the ‘Corona Crash’ case study to illustrate how to use the LPPLS method to predict a market crash. Our analysis suggests that the S&P 500 index crash in February 2020 was endogenous in nature in the sense that the market had matured into a critical regime over the previous few years, characterized by a large susceptibility to external shows and the likelihood of a burst. In this regime, the burst did occur and was triggered by the exogenous Corona crisis shock. There are various bubbles similar to the Corona Crash, and they can be predicted in advance using the LPPLS model. The second paper applies the Event Study methodology to statistically investigate whether the ‘LPPLS Confidence Indicator’ can help predict the presence of price bubbles and crashes ex-ante and causally. The research utilizes both American and Chinese industry-level data. The empirical results suggest that the LPPLS method can identify regime shifts of both positive and negative price bubbles. Specifically, positive price bubbles contain two subclass regime shifts: (i) larger positive LPPLS Confidence Indicators can detect faster-than-exponential (super- exponential) price growth, followed by a drawdown or crash, and (ii) smaller positive LPPLS Confidence Indicators can detect faster-than-exponential (super-exponential) price growth followed by a plateau, suggesting convergence to a relatively stable price level. The first regime change occurs as a result of the existence of a price bubble. In addition, the stronger negative bubbles detected by the LPPLS Confidence Indicator values are associated with higher price volatility, which breaks the symmetry with the price pattern documented for positive bubbles. This research can assist professional investors, financial institutions, and momentum-strategy funds to detect portfolio tail risks in advance and potentially avoid the losses associated with a market crash. Chapter four explores how to use machine-learning methods to diagnose financial statement fraud (which cause fundamental bubbles). Specifically, we manually collect fraudulent cases selected by sophisticated short sellers along with standard (non-fraud) company cases in the U.S. market. We then use this dataset to train nine machine- learning algorithms to classify the fraud and non-fraud firms, based on well-known financial factors, financial variables, and accounting red flags. The results suggest that machine-learning algorithms can identify the patterns of fraudulent cases quite well with only a handful of financial statement features, indicating a potential fully automated financial analysis. Based on the results and short-selling reports, we propose the ‘Polytope Fraud Theory’, which identifies ten accounting issues that can be used as a checklist to detect financial statement fraud. We then use the famous Enron case to exemplify the Polytope Fraud Theory. In addition, we propose the ‘Unified Investor Protection framework’ (UIPF), which categorizes investor protection-related theories from macro-, middle-, and micro-levels. This framework can act as a financial education material for investors to understand the general fundamental risks at different timescales. Chapter five summarizes the ‘Bubble Triangle’ and explains price-related and fundamental-related bubble detection methods. It offers our conclusions based on the empirical results of Chapters three and four. It also discusses the limitations of our research and suggests possible future research directions