Please take a look at the following thread for a solution. alex   https://communities.sas.com/t5/forums/searchpage/tab/message?filter=labels%2Clocation&location=forum-board%3Aforecasting_econometrics&q=Hourly+Forecast+Error%3A+Proc+ESM+Interval+%3D+Hour   ... View more

PROC SIMILARITY only supports univariate similaity at this point. Multivariate similarity is currently in the developemnt phase. Thus you will have to do similarity by each variable. ... View more

You are right for both questions. Since forecast studio license includes SAS/ETS, you should be able to access PROC SIMILARITY via a regular SAS session. Also the each state should be a column in the input data for PROC SIMILARITY.   here is an example from the proc documentation:   This simple example illustrates how to compare two time sequences using similarity analysis. The following statements create an example data set that contains two time sequences of differing lengths. data test; input i y x; datalines; 1 2 3 2 4 5 3 6 3 4 7 3 5 3 3 6 8 6 7 9 3 8 3 8 9 10 . 10 11 . ; run; The following statements perform similarity analysis on the example data set: ods graphics on; proc similarity data=test out=_null_ print=all plot=all; input x; target y / measure=absdev; run; ... View more

I did a Google search on Neural Network in SAS Macros and the following SUGI paper showed up. http://people.orie.cornell.edu/davidr/or474/nn_sas_implement.pdf   The paper is rather old but you can at least run a simple verion of neural network model without SAS Enterprise Miner. The neural network model implemented in EM is much faster and have a lot more model training options. Futhermore, EM comes with Random Forests and Gradient Boosting Tree models that are commonly considered for utility forecasting. Something you might consider licensing in the future.   Before getting into too much in the so called machine learning algorithms for utility forecasting, you might consider other modeling approaches that often time out beat the machine learning algorithms. Please check out the follwoing website for more info. http://www.drhongtao.com/publications   You can also try running UCM models by each hour of the data (assuming the utility series is hourly basis) and use the CYCLE components (basically a bunch sin's and cosin's with different frequencies) to capture multiple seasonalities existing in the hourly data. Here is a sample code to run UCM by each hour for an hourly utility load data:   proc ucm data=load; id date interval=day; by hour; irregular; level; cycle period=7 rho=1 noest=(period rho); cycle period=3.5 rho=1 noest=(period rho); cycle period=2.3333 rho=1 noest=(period rho); randomreg cos1-cos16 sin1-sin16; model eload = dec24 dec25 dec26 jan1 jan6 aug15 easterSat easterSun easterMon easterTue holidays holySat holySun bridgeDay endYear; estimate back=14 plot=panel; forecast back=14 lead=14 outfor=book.loadfor plot=decomp; run;     ... View more

you would need 2 dummy variables for the percenticle values: LowX3 and MedX3 (or whichever 2 you pick). X3 with high percentile values can be represented by setting LowX3 = MedX3 = 0.   original model X1 + X2 + X3 + C + X1*X2 + X1*X3 + X2*X3 + X1*X2*X3   New model X1 + X2 + X3 + C + X1*X2 + X1*X3 + X2*X3 + X1*X2*X3  + LowX3 + MedX3 + X1*LowX3*X3 + X1*MedX3*X3 + X2*LowX3*X3 + X2*MedX3*X3 + X1*X2*X3*LowX3 + X1*X2*X3*MedX3     ... View more

I think you should include X3 in addition to the percentile dummies. You can test X3 to see if it is significant after inlcuding the percentile dummies. Also interations are possible predictors to be considered if data shows different slop of X3 for each percentile value. ... View more

I can think of 2 approaches:   1. "Making the history". Basically copy the data of the current year to last year so that the hierarchical aggregation won't be affected by the missing historical data. You can make some trend adjustment while copying the data by looking at the trend from other states. You can also try to run a regression model on all states or just the states within a region to predict/impute the missing history,   2. "Leave it out". If the data for a state is short and not reliable, leave it out so that you can get the best possible forecasts for other states. You can then some similarity analysis between the state in question and other states to find a similar states using the existing 13+ month data. You can use PROC SIMILARITY (SAS/ETS) to measure the similarities between time series. Once you have identified the similar states, you can average out the forecasts by each period and then adjust the level based on the 13+ month of data. Basically you are borrowing the shape of the forecasts from similar states and adjust the forecast level based on the state's own data.   Hope this helps   ... View more