7-Day Forecast — Spend, Calls & Revenue Outlook
Forecast model: ridge regression on temperature, precipitation, snow, wind, weather code, day-of-week, and month — trained on your historical SAMEDAY daily totals. Baseline = average of same day-of-week × month.
Model Accuracy — Walk-Forward Backtest
For each day in the last 30 days we re-trained the model using only data from before that day, then predicted that day's metrics out-of-sample. Lower MAPE = more accurate.
Backtest forecast vs actual for the selected metric (use the toggle above):
Show daily backtest table (forecast vs actual + error)
Heads Up — Anomaly Alerts (with drivers)
Last 60 Days vs Forecast
All forecast bars/lines below show expected outcomes if nothing changes — i.e. you keep current bids, budgets, and tROAS targets. They are not recommendations. See the Spend Sensitivity section below to see what changes if you push spend up or down.
Daily Cost (PPC + LSA) vs Temperature
Daily Calls vs Temperature
Revenue (paid) — Actual + Forecast
Calls by Weather Bucket (historical avg)
Spend Sensitivity — What if you push PPC up or down?
Per-day what-if scenarios. Highlighted scenario maximizes paid-revenue ROAS for that day. Recommendation considers diminishing returns + current efficiency.
How this works (assumptions & caveats)
We fit a power curve to your history:
calls = a · ppc_costβcalls and the same shape for revenue. β < 1 means diminishing returns — typical for paid search.
Each forecast day starts at the model's expected spend (your "do nothing" number). We then scale that up/down and re-scale calls + revenue using the historical elasticities.
Caveats: (1) elasticities are pooled across all weather/seasonality, so on extreme weather days the response may differ, (2) the model assumes Google can spend the higher budget — at very high lifts auctions may saturate, (3) LSA spend is held constant in scenarios (it's not bid-managed the same way).
Comparable Historical Days (per forecast day)
Note: the regression model is trained on your full ~800-day history — comparable days are a separate "look-alike" sanity check. For each upcoming day we find the 10 closest historical matches by temperature, precipitation, snow and weather code. The averages from these are shown alongside the regression forecast so you can ground-truth the model with real days that "looked like this."
Recommendations for the Week
Raw Daily Aggregation (debug)