I use Vector Autoregressive model to forecast multi-assets. I use 1-lag model for example.What I have found is, sometimes the predicted value is extremly close to the actual value, the two curve almost overlap, correlation over 95%. But sometimes the result is not satisfactory, and it's very easy to find the lag between predicted and actual value. It's just like using yesterday's price as a prediction for today' price. I wonder why there is such signifcant difference between forecasting results. And I found usually we go to first case when the magnitude of the data is large, e.g. 2. 3..., and go to the second case when they are small, e.g. 0.003.I guess that's because when magnitude is large, small difference is not that obvious relatively speaking, but it is when the magnitude is small. e.g. for 2.... 0.01 is not a big deal, but for 0.003, even 0.001 is significant. Does that make sense?

maybe it's similar to the fact that stopped clocks show precise time twice a day

Did you run the VAR with levels or returns? Did you have any causal relationships in mind, or is this just data mining? If "yes" to either (especially the first), be very wary. VAR is basically (in estimation) multiple OLS. Thus, if you run it with nonstationary series (like, say, asset prices), you're going to get spurious correlations left & right. There could be other issues, but this is a good start to check.

hi wolf87, thanks for your advice.sth to make clear. do you mean it's better to run data with same levels? My data consists of 6 columns, 3 of them are 0.001 level, 2 of them about 0.1 level, 1 of them 1 level. Acutually I don't have idea what they are, maybe some are returns, some are index, such like GDP... As I mentioned, the prediction value of those 1 level are most precise--usually e.g. actual value is 3.6, and I get 3.4 for prediction. From graph, two curves are almost overlap. But for those 0.001 level, not good, even got different sign.I tried to run VAR(ECM) for three 0.001 level data such that they are same level. but still got unsatisfactory result. It's very wield that it doesn't work well for return but does for price....If I recall right, ECM can deal with non-stationary series, and I tried both VAR and ECM for data.Any ideal is appreciated!