Forecast the revenue

[yongqiangzhangzazzle]

This is a follow up question to this post: #267 (comment)

1. From OptimizationResults, assume the spend data is at the daily basis, can we get the revenue forecasting at the daily level?
2. Does the forecasted revenue include the baseline revenue? If not, how to get the baseline revenue?
3. I was trying to get the prediction results in a different way, but not sure if that is valid. We can append the spend data from the same quarter in the last year, set the spend budget for the appended data, and make the appended data as holdout data. Then once the model is trained, the prediction is applied to all the data (including the appended data), then we can get the the prediction results at the daily level with the following codes:
   meridian = model.load_mmm(model_path)
   model_fit = visualizer.ModelFit(meridian)
   df_prediction = pd.DataFrame(model_fit._model_fit_data.expected[0, :, 0])
   The problem with this approach is that the predicted revenue for the appended data is quite different from the revenue in the last year, even if the spend is the same.

Thoughts?

[cpulavarthi]

Hello @yongqiangzhangzazzle,

Thank you for contacting us!

1. I would first like to mention that Meridian currently doesn’t support predictions such as forecasting revenue or budget optimization for media spend in future quarters. However, there are some workarounds to obtain those results in order to get an idea of what to expect in the future. The comment by @alexwakim that you referred to (linked here) shares one such method to plan future budget allocation using data from a previous time period.
   Additionally, please also note that though Meridian is agnostic to the time granularity (daily, weekly, monthly, etc), daily data can be noisy and likely has within-week seasonality which may require high knots value to capture (and may potentially overfit). Aggregating to the weekly level would help resolve this issue. Please check our documentation for recommendations on Input Data Granularity for more details.

2. Your method to find the revenue forecasts can give you approximate estimates of what can be expected in the near future. The difference between the predictions and the actual revenue can be attributed to differences in factors such as reach, frequency, etc., between the training and holdout periods, though the spend is correctly accounted for.
   Meridian uses Bayesian Causal Inference and its main goal is "Causal Inference". That is, to infer the incremental effect that treatment variables have on the KPI. Without requiring to forecast the baseline into the future, users can optimize the impact of the incremental effect and understand the incremental effect.
   Meridian does not forecast the baseline into the future, because it’s not needed for the causal inference goals. Forecasting the baseline involves predicting future values of control variables and extrapolating the time series modeling beyond the time window of the provided data.

3. Holdout sets in Meridian are typically expected to be randomly chosen from the training dataset (Ref - Holdout Observations). Hence, using this for forecasting is expected to give incorrect estimates. I would recommend minimizing the forecast period to get more accurate predictions.
   I would like to reiterate that the goal of Meridian is causal inference of the treatment variables, and not forecasting into the future. Meridian needs data near knots in order to estimate them. Holding out large contiguous time chunks can lead to the knot values being driven more by the prior than by data.

Feel free to reach out if you have any further queries.

Thank you,

Google Meridian Support Team

[alexwakim]

Meridian uses Bayesian Causal Inference and its main goal is "Causal Inference". That is, to infer the incremental effect that treatment variables have on the KPI. Without requiring to forecast the baseline into the future, users can optimize the impact of the incremental effect and understand the incremental effect.

I would like to go into a bit more detail around this point. Meridian does indeed estimate the baseline during the timeframe for which it’s provided data. It does this in order to determine the difference between the actual revenue with the baseline. This is called the causal effect or incremental outcome of the treatment variables. Meridian does not need to forecast the baseline or revenue into the future in order for its causal inference to be useful for future planning. In fact, Meridian has methods to help with future planning including our Optimizer class as well as the new_data argument in many methods including roi, marginal_roi, and incremental_outcome. The new_data argument allows one to use Meridian’s causal inference to estimate quantities for any hypothetical media execution or flighting pattern.

Please refer to our glossary for definitions of some of the terms I used above.