I have data set on several SKU (within one Brand, which were divided by 3 groups) daily demand and prices during one month. Prices were fixed in this period. After this period we begin to increase prices in Group 1 by 1% (on weekdays), stay constant in group 2, and decrease in Group 3 by 1% (on weekdays). We also have demand and price data during test period. How should I analyze price elasticity of brand demand? Which model perfect fit for this data? All analysis should be understandable and can be done using Excel. Thanks in advance.
Personally, I wouldn't go down the Box-Jenkins road. Transfer functions are really pretty limited in the number of predictors and a large amount of time series data is required just to initialize the models.
Pooled time series approaches (aka event modeling in the social sciences or mix modeling in marketing) are quite flexible and conservative wrt the amount of information required. One reference providing a good, general overview is free and downloadable from Lee Cooper's UCLA website and titled Market Share Analysis. Cooper's framework is marketing but price elasticities are given explicit consideration, the prescriptions are generalizable to any discipline and he provides specific examples of applied data structures based on supermarket scanner data to leverage in answering the questions you've posed.
Another, more academically technical reference is Wooldridge's Econometric Analysis of Cross Section and Panel Data. Wooldridge has much less to say about pricing than Cooper, whose recommendations regarding price elasticities are extensive.
I would myself use ARIMAX or transfer function modeling where aggregated demand series is explained via price index.
You could form an price index which is normalized to 100 for each individual group in the base period and then chained by the daily change. Aggregated Laspeyres type on price index is obtained via weighting individual index series by the value shares of groups inside the Brand in the base period.
Transfer Function modeling is not limited by the number of predictors and identification does not require large amounts of data. Identification is based upon the relative importance of lags and is quite easy when the signal is clear. After accounting for omitted deterministic structure like pulses/level shifts/local time trends, identification is often straightforward. I have been able to automatically identify structures and suggest that you review AUTOBOX which I have helped develop. @Analyst is quite correct in his suggestion. Please see http://www.autobox.com/cms/index.php/blog/entry/elasticities-for-all for efficient elasticity computation strategies.