How do you forecast GSec rates?

Why do i care? Because my boss asked me for a simple model which predicts 10 year yields. Why do you care? Since you are reading my blog, may be you have some interest in the random thoughts running through my brain. So bear with me.

I did what everyone does. Google. The paper written by my Economics prof Rudra (say Rudro, he is very particular about the pronunciation) takes into all sorts of macroeconomic parameters, takes the data from a particular time period, generates the important factors that affect 10 year yields, creates a linear equation with a lag, makes an out of sample prediction for a period of 8 months (8 data points), calcualtes the RMPSE (root mean predictive squared error and compares that with RMPSE with a model generated with a trend line and says what the improvement is.

Cool. Next I asked our economist team here if they have done any econometric forecasting like that (I havent taken any econometric courses to replicate the method). Not only have they not done anything like that, but the guy says the forecasting of long term yields is rubbish, as the rates are entirely based on expectations and not based on history. In any case, the volatilty is unprecedented.

It is a universally acknoledged fact that NOW is always somewhat unprecedented. But as long as you take the right factors, you should still be closer to the truth. After all, expectations do not come from the heavens, but based on the data available. I would use my limited knowledge. All I want is a simple model.

I took exactly the same data (BSE100, REER, M3, WPI, Oil prices, IIP etc.) I dont have the tools to account for multicollinearity (two X factors are correlated) but I can still filter the factors based on t-statistics till all the factors in the regression equation have low p-values. Also, forecasting requires that I take all the inputs with a lag.

After 2 iterations, I filter down to just 3 factors - REER(t-1), IIP (t-1) and GSec (t-1). Based on the out of sample forecasts the actual and forecasted came close (forecast period Mar to Oct 08).

To be fair, the last month was unprecedented. The yield fell practically over 200 bps (from 7% in November end to low of 4.86% in December end) and then rose 120 bps last week. The fall of 200 bps was expected; The additional government borrowing of Rs. 5000 crores which pushed the yield up 120 bps was unexpected. Any model however correct (I asked the economist to check and if possible refine it) has to be supplemented with Qualitative data. Inspite of this, and the recent discredited models of credit rating agencies and Taleb's wide popularity, I think a model gives the relationship between variables elegantly in a way no amount of theory could.