Multiple Regression Forecast I made a model based on approx 75% of the data set(in the sample). Then applied the model's estimator on the remaining values of the data set. The regression had approx 40% adj R^2.
Google lets you download 90-day interval of daily data, which has to be merged with weekly SVIs in order to have the same scale. I believe this is the reason why my prediction is so jagged compared to the actual values.
Any suggestions on how to mitigate this?
Forecast: Blue = Actual, Orange = Predicted

 A: You have time series data which is not the same as cross-sectional data as there may be trends , level shifts , dynamic response to causals etc.. OLS (simple regression) does not incorporate memory unless you introduce appropriate lags in your variables. Time series methods identify the appropriate/required lag structure and anomalies . Your tool of choice is wanting ... why don't you post the weekly data fro both series in a csv file and I will demonstrate the art of the possible. Prediction lines are often never as smooth as actual values (but can be !) but your prediction line seems biased suggesting possibly inferior analytics. Analytics that are free often come with a subsequent price.
Please be specific about your "training data set" and your forecast region including any specification of future values for the predictor series.
Google output would also help explain their potential analytical shortcomings.
EDITED AFTER RECEIPT OF DATA :
It is inconguent to me that GOOGLE would a 2 variable OLS model on time series data. OLS models assume (by specification) that there are no lags needed in a predictor series to capture the relationship between the two observed series. Furthermore there are no untreated anomalies i.e. unusual values in either of the two series. Additionally there is an assumption that neither series has been effected by an individual exogenous event (e.g. a level shift) or that both series have been similarly effected. Furthermore there is the assumption that the errors from the model have constant variance. Furthermore there is the assumption that the parameters of the model are invariant over time. I suggest taking GOOGLE's free analysis with a big grain of salt.
Now the bad news is that your data has missing days ( holidays and such ) . Time series methods (ARMAX) including causal variables require a complete accounting. Now you could consider Your series to be a 5 day per week data series WITH  holiday values to be interpolated. The 5 day-per-week series may have a day-of-the-week effects which need to be considered in order to get a good reading on the response of Y to X . This day-of-the-week effect can either be deterministic or stochastic(adaptive) . If you wish to post two columns of 5 day-of-the-week data I will see what I can do to bring clarity to your analysis.  
