OK, now that I’ve written all that, I guess I can answer better: I’m more comfortable doing analysis to improve forecasting myself, but ultimately we’re in this business to make money, and so one has to make the trade or construct the portfolio. But my instinct, when using these kinds of models is to use a shrinkage factor, chosen subjectively, to account for the fact that these models always look more certain than they really are.
LPoulin133 Wrote: ------------------------------------------------------- > maratikus - To your point, most literature focuses > on the latter (as far as ‘complete’ models go), > rather than the former, though my sense it takes a > certain level of infrastructure to prove viability > in practice. Has this been your experience? I want to make sure I understand your question. Are you talking about spread trading vs error-correction model (improved forecasts)?
maratikus Wrote: ------------------------------------------------------- > I want to make sure I understand your question. Are you talking about spread trading vs error-correction model (improved forecasts)? Yea, since you can spread anything in almost any manner with as much discretion as you want there are many ways to evaluate its effectiveness. As you say, directly benefitting from the spread itself. I guess what I’m asking about is the effectiveness in practice of ECMs on low frequency timeframes. Consistently, I’m told it isn’t, though I don’t personally know.
What is ECM again?
error-correction model (or equity capital markets, he).
Ah right right… got confused by acronym collisions
sorry guys just one more question: if my original time series is stationary, but then i transform the original time series to a Moving Average of span 2 and then run an AR(1) model on the transformed series, would this be ok even if the transformed time series is now not stationary? thanks!
It would be better to get some context, but here are some thoughts: When you say moving average of span 2, do you mean MA(2) or do you mean just a two-period moving average of the original series? If you mean MA(2), then you could just instead estimate ARMA(1,2). If you mean a two-period moving average, then I’m not sure what the purpose would be of running the AR(1) since by definition the value in t will be correlated with the value in t-1.
thanks jmh, i mean a two period moving average. basically im still trying to estimate mean reversion. so i computed weekly averages of my rates and im still getting mean reversino estimates that are way off from the expected range. however, if i smooth the data with this two period moving average, the results look much better. but i guess what you’re saying is it doesn’t really make sense to run an AR(1) model on this?
jimjohn Wrote: ------------------------------------------------------- > sorry guys just one more question: if my original > time series is stationary, but then i transform > the original time series to a Moving Average of > span 2 and then run an AR(1) model on the > transformed series, would this be ok even if the > transformed time series is now not stationary? > thanks! If the original series is stationary, then its moving average is also stationary.
bchadwick Wrote: ------------------------------------------------------- > OK, now that I’ve written all that, I guess I can > answer better: > > I’m more comfortable doing analysis to improve > forecasting myself, but ultimately we’re in this > business to make money, and so one has to make the > trade or construct the portfolio. > > But my instinct, when using these kinds of models > is to use a shrinkage factor, chosen subjectively, > to account for the fact that these models always > look more certain than they really are. It’s always wise to be cautious of over-curvefitting. Your approach seems practical.
maratikus thats what i thought too, but i tried running a Dickey Fuller unit root test on the original data and it passed, but when I ran the test on the moving averaged data it failed. or does the test only apply to the original data and im not supposed to test the moving averaged data? thanks!
That sounds correct. What does an autocorrelation plot look like? Sounds like you should at least try the Box-Jenkins method and see if an ARMA(1,2) procedure looks plausible. The MA part of things essentially smooths out the noise similarly to what you would expect with a normal moving average, but it does it in a way that is compatible with having an AR. The difference between AR and MA is basically that the AR tells you that past values of the series are useful for predicting the future values of the series. The MA part tells you that the UNEXPECTED part of past returns has its own predictive dynamic. A general mean reversion model would presumably have an AR component, because high values would head toward the mean and vice versa. But if the unexpected noise elements have an unusually strong or weak effect on mean reversion compared to an ordinary AR process, then you’d want to model that with MA.
LPoulin133 Wrote: ------------------------------------------------------- > I guess what I’m asking about is the effectiveness > in practice of ECMs on low frequency timeframes. > Consistently, I’m told it isn’t, though I don’t > personally know. I’m not sure. It really depends where you try to apply those. For example, if you think that the housing prices and CPI are co-integrated, then an ECM would give you a better forecast of both. The adjustments would be the highest when the spread is high. If I were you, I would test the effectiveness of the ECM yourself and make your own conclusions. What might work for somebody else would not necessarily work for you and vice versa.
…you’re right, I had assumed equities.
jimjohn Wrote: ------------------------------------------------------- > maratikus thats what i thought too, but i tried > running a Dickey Fuller unit root test on the > original data and it passed, but when I ran the > test on the moving averaged data it failed. or > does the test only apply to the original data and > im not supposed to test the moving averaged data? > thanks! DF test is just a test that gives you an answer with a certain confidence level. Let’s just think about it. If you have two series (the original and the lagged one) that have finite variance, how can their sum have infinite variance? That doesn’t make sense. Maybe I misunderstand what you did, or maybe you didn’t run the tests correctly or maybe the confidence level is not low enough.
bchad, im not all that familiar with analyzing the ACF and PACF functions or the Box-Jenkins method, but the ACF seems to be decreasing very slightly as the lag increases. not by much though, overall looks pretty constant. the PACF has a big spike at lag one and then dies down after. this is for my original data set. maratikus, hmm i guess that makes sense. i actually didn’t compute the moving averages manually, i used SPSS and computed what was called a “Centered Moving Average of Span 2”. i assumed this was just a two period moving average but maybe not? ill look into it. the DF confidence elvels were very different though. something like a p value of .0063 for the original and .13 for the smoothed series. thanks again guys, appreciate teh help!
jimjohn Wrote: ------------------------------------------------------- > bchad, im not all that familiar with analyzing the > ACF and PACF functions or the Box-Jenkins method, > but the ACF seems to be decreasing very slightly > as the lag increases. not by much though, overall > looks pretty constant. the PACF has a big spike at > lag one and then dies down after. this is for my > original data set. That’s huge evidence in support of AR(1) model.
I agree. In general, try to follow KISS. Don’t need to get fancy when you don’t need to.
ok guys i just tried computing the two-period moving averages manually and im still getting many cases where the original series passes the DF test but the smoothed series does not. any ideas what the issue could be here? thanks!