In my opinion, If we consider the fluctuations (events that effectively effects) in the past, and incorporate some multiple structural breaks. we can rescue forecast from overestimation.   

Dear Paul

Your question is very interesting, but it is very difficult for me.

Forecasting time series can be a very hard task due to the inherent uncertainty nature of these systems. It seems very difficult to tell whether a series is stochastic or deterministic chaotic or some combination of these states. More generally, the extent to which a non-linear deterministic process retains its properties when corrupted by noise is also unclear. The noise can affect a system in different ways even though the equations of the system remain deterministic. Since a single reliable statistical test for chaoticity is not available, combining multiple tests is a crucial aspect, especially when one is dealing with limited and noisy data sets like in economic and financial time series.

I also suggest this paper

https://www.researchgate.net/publication/271683439_Forecasting_Exchange_Rates_A_Chaos-Based_Regression_Approach

Article Forecasting Exchange Rates: A Chaos-Based Regression Approach

In my opinion, best we can rely on, if anything, is what Thomas Walther suggested. 

Well, deep subject; time series has both advantages and disadvantages. Some advantages – you can compute fast, you don't have to develop complicated models, and you can eyeball fluctuations in trends.  

Disadvantages – since basically a series explains itself it is hard to assess structural changes like you could with explanatory variables.I find them very helpful to assess trends in my stock investments, again because the method is quick and can give me a snap shot of whats going on, but to predict the future, hmmm...Now, if you can include macroeconomic variables into the equation you can improve the forecasting ability.

I agree with previous comments, Time Series have pros and cons.

I just want to add that to reduce uncertainty in the future horizon, using multivariate methods can integrate relevant information to the Time Series Forecast. This information can be incorporated in statistical methods (multiple linear regression, dynamic regression) or on soft-computing methods (such as neural networks). In addition, there are currently technologies like Big Data and Open Data that could provide interesting sources to help forecast the time series information.

May want to also consider how over time the effects of crisis on different cross-sectionals which time series alone would not able to account. Hence, to look at panel data analysis that analyzes cross-sectionals over times.

Best wishes 

Paul -

If Thomas was suggesting adding regressors that include current data, such that you really have a combination of a time series for your variable of interest and other time series information that reach what is considered in your study to be the current time period, then I think that that could help. However, you need to be careful not to include too many regressors. I have seen an overuse of lags obviously overfit a model so that it deteriorated badly overtime, even though it used time series data for at least one regressor that extended beyond that of the variable of interest. Of course, applications and experience may vary substantially.

The problem with relying basically on time series, as indicated by others, is that a time series forecast cannot tell you what is happening now, because it does not use current data. You stated that "AR, ARCH. GARCH, ARIMA, etc. do not seem to be helpful in forecasting the coming of crisis." That is because, even with consideration for such phenomena as seasonality, they cannot tell you about a coming new crisis for which they use/have no information. That is why I think that including a regressor with data relevant to the current situation may help, but might also complicate interpretation, especially the variance interpretation.

I have attached a link to some notes I wrote that I constructed to (1) explain the difference between 'prediction' for cross sectional surveys versus time series 'forecasts,' and (2) to explain why a strictly historically based time series is not generally a good imputation method for a cross sectional survey.

I warned in those notes that cross sectional 'prediction' and time series 'forecasting' should not be mixed. Now that Thomas' answer reminded me of adding regressors that may have time series information that extends to what might be considered the 'current situation' - I would consider that possibly worth exploring, but it depends on your application. I have seldom seen it come up. And the basic idea of depending upon time series, rather than regressing current sample data with regressor data on the whole population, judiciously applied one similar stratum/group per model, is favorable to a cross sectional approach performing best, when you want to know the current situation.

All of this argues that a current crisis needs a cross sectional survey, but what about future crises? Well, because we have no data on future 'surprises,' there is no way to tell. Generally the further out into the future that we forecast, the greater the inaccuracy, not just from variance when we assume the model is correct, but because of those occasional breaks that will come in all time series data, as you indicated. If you have some knowledge of possible future changes, perhaps an invention that may change a market, or a treaty that restricts some product, you could do some subjective sensitivity analyses. You might be able to rerun your time series model with some changes to see "what-if," and some analysts will produce graphics with high and low options. But when you do not have any data for the period of interest, you do not know when something changes.

Cheers - Jim

Research When Prediction is Not Time Series Forecasting

All models are false, but some are useful.

Crises are, by definition, outlying events, which are unpredictable.

No model is successful in predicting crises.

https://arxiv.org/ftp/arxiv/papers/1302/1302.6613.pdf

I think, it's the pre-assumed linear form of these models. While generally Most real-world time series are basically non-linear