Forums  > General  > merging forecasts for varied time horizons?  
Page 1 of 1
Display using:  


Total Posts: 1540
Joined: Jun 2004
Posted: 2019-05-12 19:09
Here is a problem and, as usual, I am being stupid.

I am trying to predict an overnight interest rate in one of the emerging markets in several days (up to 5 weeks). There is a relatively large history of observations (say 5-6 thousand observations) and I found a few predictive models.

Model 1: reversion to the long-term mean - produces a forecast for time now + n

Model 2: short-term micro-reversion- produces forecast for tomorrow (now + 1)

Model 3: day of week seasonality - produces adjustment based on the day of the week

What I am trying to do now is merge these models into a single forecasting method. The problem is that (a) the models are not strictly independent and (b) I want to understand if I am getting an improvement (the base line is unchanged from todays rate).

I don't interest myself in 'why?'. I think more often in terms of 'when?'...sometimes 'where?'. And always how much?'


Total Posts: 1260
Joined: Feb 2005
Posted: 2019-05-13 11:43
wouldn't you just use some form of ensemble learning to combine the forecast?

Dilbert: Why does it seem as though I am the only honest guy on earth? Dogbert: Your type tends not to reproduce.


Total Posts: 446
Joined: May 2006
Posted: 2019-05-13 12:14

d Prediction = ShortTimeScale * (ShortTerm - Prediction) dt + dSeasonality + dW
d ShortTerm = LongTimeScale * (LongTerm - ShortTerm) dt + dZ


"There is a SIX am?" -- Arthur


Total Posts: 679
Joined: Jun 2005
Posted: 2019-05-13 14:41
Minimize Sum of error normalized by combined error prediction ~ R2.

pred_i (i=1,2) depends on model parameters.

R21 ~ abs(pred1-price)/err1
R22 ~ abs(pred2-price)/err2

R2 ~ [(pred1-price)^2+(pred2-price)^2]/[err1^2+err2^2]

R2 behaves well, i.e. it roughly follows Chi2.

You also can apply inverse-CDF transformations (it's also called PIT) and combine probabilities (using ML/NN?). Usually I check something like:

if 2*p-1 \in [-1,1]
then (p1+p2-1) \in [-1,1]

or something like...

PS. typos..


Total Posts: 22
Joined: Feb 2018
Posted: 2019-05-13 16:09
Could Fernholz with his hedged strategy from "Statistics of statistical arbitrage" be of any help?
Previous Thread :: Next Thread 
Page 1 of 1