Dealing with extreme shocks like global recession in time series

Question

I'm following Rob Hyndman's forecasting otext to practice on some financial data for fun and I am having difficulties in trying to properly deal with large shocks similar to the 2008 recession.

My data is as follows:

library(forecast)
stock <- c(60,150,149,66,28,59,75,148,11,188,113,313,340,362,321,360,334,338,424,377,399,365,
       433,471,378,335,332,264,291,247,352,267,376,258,295,349,335,230,261,339,256,405,377,
       392,372,272,462,395,399,487,627,607,462,573,775,776,923,764,883,831,792,860,849,223,
       177,99,53,121,102,103,44,134,217,116,167,221,90,288,305,290,292,345,384,302,348,292,
       374,489,448,520,424,447,451,472,481,463,446,470,387,538,381,353,259,374,418,331,354,
       418,416,348,309,353,279,341,291,255,278,302,182,256,277,299,269,276,310,368,263,329,
       211,249,318,293,278,265,221,260,194,260,282,250,231,131,188,154,186,279,318,369,353,
       322,349,301,366,364,455,450)

stock_price <- data.frame(price = stock)
stock_ts <- ts(stock_price, start=c(1994, 1), end=c(2006,12), frequency=12)
stock_ts_train <- window(stock_ts, start=1994, 2006-.01)
fit.holt.exp_dp <- holt(stock_ts_train, damped = TRUE, exponential = TRUE)
plot(fit.holt.exp_dp)
lines(stock_ts, lty=2)

I can somewhat make out a cycle that occurs every 6 years with no observable seasonality, but this damped exponential smoothing method is forecasting a flatted trend line. My issue is that the prediction intervals seem to be too large, mainly due to the drop in 1999, and need a way to curb that influence. One simple method I was pondering was to impute the data during this large climb and drop by using a moving average and run the model that way.

I'd really appreciate any thoughts or guidance!

Thanks

Answer 1

I do not believe that you can predict regime change and that is what is happening in 2008. There is nothing in the model that stands as a parameter for this differential shift in a model that just uses time-series procedures to capture the past.

I discuss the underlying problem in this piece

https://www.researchgate.net/publication/256801634_The_practice_of_quantitative_methods

which distinguishes between open and closed systems defined in a realist sense. It is possible with a closed system - a simple machine, the clockwork universe, an experiment in which closure is achieved by design - to get unchanging regularities that make for successful predictions. In open systems the system is itself capable of changing so the gearing ratios of the machine changes and predictions are now much more difficult. Human systems - here the economy - are capable of self-change so successful prediction is hard.

I remember as a beginning postgrad many years ago a seminar debate between Fred Smith who taught Box-Jenkins time series analysis at Southampton and a visiting speaker - Oliver D. Anderson of the Civil Service College. The former expected a time -series approach should be capable of forecasting real change - the latter did not. I side with Anderson as the procedures are black box and given little insight to what is happening in a changing world.