I can’t forecast every up and down in every month and year, but I essentially know unemployment statistics will be terrible over the next decade(worse 10 year period since the Great Depression), which is easier to forecast than the short and medium terms. Are you saying you have seen and know all? And why would I take anyone’s anecdotal experience seriously.
Let me be clearer about what I mean. Evaluating forecasts formally requires us to choose a metric. In the academic literatures in economics, finance and computer science, the overwhelming majority of people opt for the expect square loss or the expected absolute loss. Usually, we also consider that our best approximation of that requires us to backtest our forecasting methods, collect the resulting pseudo-out-of-sample forecasting errors and estimate the performance criteria using an empirical average of the appropriate transformation of those errors. Hence, when I talk about "good" and "bad" forecast, I refer to those metrics either as an absolute (i.e., the number itself for a given forecasting methods) or in relative terms (i.e., comparing two or more methods).
If you look at the unemployment rate in the US, most people would choose to impose a unit root and to model the first difference. In that case, your forecast would look like this: UR(T+h) = f(data(T)) + UR(T). One sense in which you can say that something is hard to forecast is that even a relatively bad, though simple model is hard to beat. Specifically, here, an AR(1) on the first difference is relatively hard to beat: UR(T+h) = a + b*(UR(T) - UR(T-1)) + UR(T). However, if you allow yourself to look up a mass of data (say, the FRED-MD data base), you can do something about it. For example, there seems to be a sweet spot some 9-12 months ahead where nonlinear methods such as Kernel Ridge Regression and Random Forest applied on lagged changes in unemployment rate and PCA-extracted factors and their lags will do very well. On some US macroeconomic variables, including the unemployment rate, you can extend that sweet spot to include 24 months ahead.
How do I know that? Well, I read dozens of papers on the matter and those specific results come from backtests I ran with some of my co-authors. It's fair to point out that I didn't read everything there is to read and I don't everything there is to know about forecasting. I'm not sure it's entirely fair to call this "anecdotal experience" on the other hand.
The longer term is easier to forecast than the shorter in terms of percentage, so we’re already on conflicting ground.
It depends on what you mean by "easier" to forecast. There is the idea of being close to the realization on an out-of-sample basis -- so, a small root mean squared error or a small mean absolute error in a backtest. Or, to paraphrase one of my superviser, there is the idea of beating a monkey model -- i.e., trying to beat simple methods (historical means, random walks, AR(1)). In the first case, ask yourself what kind of data generating process would not imply that the irreducible uncertainty of your forecast grows with the forecast horizon... Usually, you fall further off the track on average when you ove the target further away in time. For the second sense, it depends on the variable. What we found is that there seems to be a sweet spot for forecasting a macroeconomic variable and that it depends on the variable in question. It's easier to beat simple models further in time for the UR, but the window is closer to 1-3 months if you talk about certain spreads on interest rates.
Depending on what you mean by easier to forecast and on how you evaluate forecasts, you might be correct. However, if you pick the definitions I gave above, what I said would be correct up to my best knowledge.
Have you even attempted to understand EW yet?
I did not attempt to understand EW seriously. The little effort I put landed me on absurdly vague descriptions.
On the other hand, you told me before that you have been thinking about this for years. Why not open up a new thread in the economics subforum here where you explain the idea? I am sure you can give a good summary and an example or two within a one or two posts and it would help people understand what you are talking about all the time.
