8,487 reputation
11346
bio website wiki.scn.sap.com/wiki/pages/…
location Switzerland
age 39
visits member for 4 years, 7 months
seen 11 hours ago

During the day, I create software to forecast retail sales and calculate order proposals. Lots of time series, with an emphasis on fast, automatic and robust data cleansing and forecasting. I'm an elected director for the International Institute of Forecasters and an Associate Editor for their practitioner-oriented forecasting journal Foresight. At night, I switch hats and do inferential statistics for academic clinical and biological psychology. My main tool is the statistical computing environment R.

profile for Stephan Kolassa on Stack Exchange, a network of free, community-driven Q&A sites


14h
revised Missing data paired samples t-test
added missing-data tag
15h
revised Multicollinearity in polynomial regression
formatting & multicollinearity tag
15h
answered Multicollinearity in polynomial regression
18h
comment Ensemble time series model
@denis: I'm not familiar with forward or backward predictors, nor with Burg's method, sorry... although I would assume that combining forecasts/predictions (aka ensemble methods) will usually be beneficial.
19h
comment Drawing n intervals uniformly randomly, probability that at least one interval overlaps with all others
You can look at the probability that the last drawn $A_n$ is smaller than the minimum of all previously drawn $A$, and the probability that the last $B_n$ is greater than the maximum of all previously drawn $B$. This should be helpful. Then inflate the probability to account for the fact that we don't need the last one, but any one. (I don't have the time to work through it, but it looks like a fun little problem. Good luck!)
20h
answered Using Diebold-Mariano test to compare predictive errors in non-time-series?
20h
answered R Time Series Analysis forecast result always remains same
Apr
17
comment ANOVA with single degrees of freedom
No problem. 1 df = 1 parameter. If you have two groups, one of them will be the reference. And the (single) parameter will estimate the difference in DV between the reference group and the other ("non-reference") group. Generally, if you have $k$ factor levels, this takes $k-1$ df.
Apr
17
revised ANOVA with single degrees of freedom
deleted 16 characters in body
Apr
17
answered ANOVA with single degrees of freedom
Apr
17
comment Statistical test to compare paired scores
You may be looking for a paired t test.
Apr
17
comment Assessing stability of a method to less data (subsets of it)
I'd then present the variation of your parameter estimates, e.g., by giving 2.5% and 97.5% quantiles. What the significance is will depend on what your null hypothesis is.
Apr
17
comment Assessing stability of a method to less data (subsets of it)
I'm not entirely sure I understand your question (how can you assess the weight of an object based on part of it?), but you sound like you might be interested in bootstrapping. Yes, resampling is (somewhat) random, but unless you have a really hard problem, you can do enough (>=10,000) samples to get stable results, and randomization procedures have been a standard and accepted part of statistics for at least 15 years now. (You can always do five independent bootstraps and see whether the results vary.)
Apr
17
comment Predict time series data from another
Why would linear regression not be suitable for prediction? I edited my answer to also include ARIMAX, which I really don't think is justified if you have only six observations.
Apr
17
revised Predict time series data from another
added ARIMAX model
Apr
17
answered Predict time series data from another
Apr
17
awarded  regression
Apr
16
comment ARX model selection
How do you deal with the long strings of zeros in your data? ARX will try to fit lots of consecutive zeros, which does not make sense. If you just throw the zeros out, ARX will believe that an observation three months back is just one week old, which likely is also not what you want. I suspect your first problem is "fitting an ARX model with missing values".
Apr
16
comment How is the error calculated in ETS using R?
Well, they are the residuals. $e_t=y_t-\hat{y}_t$, where $y_t$ are your observed actuals, and $\hat{y}_t$ are your ETS in-sample fit. (Possibly with $y_t$ and $\hat{y}_t$ reversed, there are various schools of thought on how residuals are defined.) I assume you are not asking how to re-implement ets() in Excel, because that would be a somewhat large project for this site.
Apr
16
comment ARX model selection
Hi, and welcome to CrossValidated! Could you please edit your question to indicate what time granularity you have (daily? monthly? hourly?) and how you know that your $AIC_c$-selected orders are "clearly overfitting"?