creating error bands

[TomDoan]

So you get reasonable results if you use seasonally adjusted data and unreasonable results if you try to seasonally adjust data yourself. Do you know how complicated seasonal adjustment is? For things like price series, the seasonality in components are quite different (heating oil has a different seasonal than gasoline has a different seasonal than electric power, etc.). The whole process of creating a price index is extremely complicated and that's even more so with seasonally adjusted price indices.

[TomDoan]

Also 12 lags have been given as an option for VAR lag selection (quarterly data) and this is generally the procedure in most papers (bank of england also suggests this).

The fact that other people do it doesn't make it a good idea. I'm not even sure I would do 12 lags with quarterly data with a prior. With OLS estimates, not a chance. And if you aren't going to pick 12 lags, why even allow for it?---you're just throwing away data points in the lag selection process.

And it generally recommends 1 to 2 lags (AIC).

1 isn't enough. Period. It doesn't matter what AIC or SBC says. (SBC often comes up with 1 lag). Almost all your series are non-stationary. The price variables may be borderline I(2). With just one lag, the only thing the model can explain is the unit root properties---there isn't enough freedom to do anything shorter term. Ideally, you would probably prefer to do 4 and you probably have enough data to do that.

And no macroeconomic data can have more than 90-100 data points. so this is a general issue widely discussed in literature.

That's nonsense.

You can see my data set begins at 1984 and we have to exclude data beyond 2007 because it gives weird results (as you suggested).

It starts in 1984 probably because you're using a series which doesn't exist prior to that. There's plenty of US macro data dating back to 1948, and quite a few papers by very well-known authors have started in 1948, or perhaps mid 1950's (to avoid the Treasury Accord period). And the problem with your original data set was that you stopped it at a very volatile period. Add more data (so that it's clear that the wild behavior at the end doesn't persist) and you're likely to go back to getting reasonable results.