Improving the accuracy of bubble date estimators under time-varying volatility

summary

The gist

In this study, we propose an algorithm for estimating bubble dates under time-varying volatility that improves accuracy compared to methods based on homoskedasticity <ref:2306.02977#pg2>.

In short

The study proposes a two-step algorithm to estimate bubble dates when volatility changes over time. First, it estimates initial break dates assuming constant volatility and then uses Weighted Least Squares (WLS) on residuals after estimating the time-varying error variance. This method significantly improves the accuracy of bubble date estimations compared to simpler methods like Ordinary Least Squares.

Key concepts

Four-regime bubble model
This model describes a market's behavior as switching between four distinct phases: a stable initial phase, an explosive phase (the bubble), a collapsing phase, and finally, a normal market phase. It uses structural changes to define these transitions.
Time-varying volatility
This means the level of market uncertainty or risk is not constant; it changes over time. The paper assumes this change is modeled by four distinct regimes, which requires specialized estimation techniques beyond standard models that assume constant risk.
WLS-based estimation
Weighted Least Squares (WLS) is a statistical method used to find the best fit for data when some observations are more reliable than others. In this context, it is used in the second step of the algorithm to estimate the bubble dates by weighting residuals based on an estimated time-varying volatility.
Sample splitting approach
This is a non-parametric technique used to estimate time-varying parameters, like volatility, by dividing the data into different segments. By estimating parameters within these segments, researchers can create a flexible model for how the error variance changes over time.

Terminology used across episodes

This episode discusses

The paper

Improving the accuracy of bubble date estimators under time-varying volatility · Read on arXiv

Eiji Kurozumia, Anton Skrobotov

Hitotsubashi University · Russian Presidential Academy of National Economy and Public Administration · Saint Petersburg University (Center for Econometrics and Business Analytics)

In this study, we consider a four-regime bubble model under the assumption of time-varying volatility and propose the algorithm of estimating the break dates with volatility correction: First, we estimate the emerging date of the explosive bubble, its collapsing date, and the recovering date to the normal market under assumption of homoskedasticity; second, we collect the residuals and then employ the WLS-based estimation of the bubble dates. We demonstrate by Monte Carlo simulations that the accuracy of the break dates estimators improve significantly by this two-step procedure in some cases compared to those based on the OLS method and to the sign-based dating procedure of Harvey, Leybourne and Zu (2020).

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Improving the accuracy of bubble date estimators under time-varying volatility".

Warren: Here's the gist of the paper. In this study, we propose an algorithm for estimating bubble dates under time-varying volatility that improves accuracy compared to methods based on homoskedasticity <ref:2306.02977#pg2>.

Ruth: The main finding is that a two-step procedure involving estimating break dates and then employing WLS-based estimation of the bubble dates significantly improves the accuracy of the estimators by Monte Carlo simulations <ref:2306.02977#pg9>.

Warren: Model Formulation The paper considers a four-regime bubble model under time-varying volatility, defined by a structural change model with break points given by ke, kc, and kr <ref:2306.02977#pg3>.

Ruth: The process is modeled as: 1.

Paper summary: Ruth: So we're looking at this paper, "Improving the accuracy of bubble date estimators under time-varying volatility." It seems like they're tackling a problem where the market isn't just moving randomly, but the way it moves changes over time, which is called non-stationary volatility.

Warren: Right. So what’s their main idea here? They propose an algorithm for finding those bubble dates—the emerging date, the collapsing date, and the recovery date—and they claim this two-step approach improves accuracy compared to older methods.

Iris: The core of it seems to be a two-step procedure that starts by estimating the break dates assuming things are constant, like homoskedasticity, and then using those initial estimates to correct for the actual time-varying volatility later on.

Ruth: Exactly. They do this first by estimating the bubble dates using a method based on assumptions of constant variance, which is what they call non-stationary volatility accommodation <ref:2306.02977#pg1>, and then they take the residuals from that fit to do something else.

Warren: And that second part involves collecting those residuals and then using a weighted least squares approach to estimate the actual bubble dates, which is what they call the WLS-based estimation <ref:2306.02977#pg1>.

Iris: What’s interesting is how they test it. They use Monte Carlo simulations to show that this two-step procedure significantly improves the accuracy of the break dates estimators in some cases when compared to methods based on ordinary least squares, or even those based on sign-based dating from Harvey, Leybourne and Zu <ref:2306.02977#pg1>.

Ruth: It matters because if we're dealing with real financial data, which is often noisy and changes its volatility depending on what’s happening in the market, having a correction method that adjusts for that changing volatility could mean finding those bubble dates more reliably.

Warren: I wonder how much this correction actually saves us in practice. The paper mentions performance gains when there's just one time break in volatility, especially if that break happens right at the start or the end of the sample <ref:2306.02977#pg3>.

Iris: That suggests the correction method is most effective when it can identify that specific change in volatility structure, which is a good constraint to keep in mind for practical application.

Ruth: So, they use this whole four-regime bubble model framework—unit root process, an explosive regime, a collapsing regime, and then the normal market—to set up this estimation <ref:2306.02977#pg1>.

Warren: And the way they handle the volatility itself is through an adaptive weight selection process using kernel-based methods to estimate that time-varying error variance <ref:2306.02977#pg1>.

Iris: That adaptive weighting involves estimating some parameters, like tau e, tau c, and tau r, by first using a sample splitting approach based on ordinary least squares with a delta t of one for all time points <ref:2306.02977#pg2>.

Ruth: Then they use the least squares method to get residuals and then plug those into the variance calculation using a bounded nonnegative continuous kernel function, K(s) <ref:2306.02977#pg2>.

Warren: That sounds computationally intensive. It’s not just plugging in numbers; it’s running an estimation process that calculates the error variance based on where you are in the data, which is what they mean by "the correction depends on where you are" <ref:2306.02977#pg1>.

Iris: From an econometric standpoint, the identification here relies on using those estimated break dates as inputs to estimate the volatility function non-parametrically <ref:2306.02977#pg3>. The question is whether that identification holds up when you move beyond this specific four-regime setup.

Ruth: It does, based on their Monte Carlo results showing accuracy at least as good as the homoskedasticity assumption and better in some cases <ref:2306.02977#pg1>. It's a solid improvement over what you get with just the OLS method <ref:2306.02977#pg1>.

Warren: The empirical application using different time series of cryptocurrencies really backs this up, showing that the WLS method returns adequate break dates more often than the simple OLS approach <ref:2306.02977#pg1>.

Iris: So, to summarize, they propose estimating dates without volatility correction first, then using a WLS-based estimation after non-parametrically estimating the time-varying error variance from those residuals. It’s a structured way to handle the changing market environment <ref:2306.02977#pg3>.

Ruth: And the conclusion is that this correction method gives estimated break dates that are at least as accurate as if you assumed constant volatility, and sometimes better, depending on where the volatility change happens <ref:2306.02977#pg1>.

Warren: It means for anyone looking at bubble detection in financial markets, especially when dealing with data like crypto where things are volatile, this two-step procedure offers a more robust way to find those critical dates <ref:2306.02977#pg1>.

Conclusion: Ruth: So, we’re wrapping up on this paper about improving bubble date estimators under time-varying volatility.

Warren: Basically, they took a four-regime bubble model and proposed a two-step way to find the emerging, collapsing, and recovering dates that should be way more accurate than the old methods.

Iris: And it does use an initial estimation based on constant volatility assumptions first, then corrects those dates using a weighted least squares approach after looking at the residuals.

Ruth: So what does this actually mean for anyone tracking financial bubbles?

Warren: It means if you’re using a model that assumes the market volatility stays steady, you might be off by a lot, but this method tries to account for when that volatility actually changes.

Iris: The authors show in their simulations that this two-step correction process gets much closer to the real dates than just using ordinary least squares or even some sign-based dating methods.

Ruth: That’s the big picture, right? It suggests that ignoring how volatile things get is a major source of error in bubble detection.

Warren: Yeah, and they found that this correction actually works better when there's a single time point where the volatility structure shifts, especially if it happens at the very beginning or end of the data you’re looking at.

Iris: That specific condition is important because it tells us exactly what kind of market event this method is best at catching.

Ruth: It really brings the accuracy up to the level of what you'd get even if you just assumed constant volatility, but better in those specific cases where things change.

Warren: So, for someone just listening who isn't deep in econometrics, it boils down to this: using a two-step correction gives you better timing on when bubbles start and end.

Iris: And the authors confirm that this approach is empirically effective when tested on real cryptocurrency data, showing that the weighted least squares method finds these dates more often than the simpler ordinary least squares method.

Ruth: It’s interesting to see how they combine those different estimation techniques—the initial break date estimates and then the non-parametric variance estimation—to get that final result.

Warren: The main thing is that it gives us a better tool for timing things in markets where the underlying rules of movement are constantly shifting.

Iris: But remember, this method relies on them correctly identifying those initial break dates first, which is where the identification really rests.

Ruth: Next up, we're going to look at what they actually did to figure out that time-varying volatility function itself.

More episodes

← Home