Foundation AI helps Electrolux optimize their marketing budget using Advanced Analytics

Extract data - foundation AI

Foundation AI is an Artificial Intelligence Solutions Provider. We help organizations process, manage, and leverage their unstructured data to automate labor-intensive tasks, make better data-driven decisions, and drive real business value.

Electrolux - Extract Data

Electrolux is a Swedish multinational home appliance manufacturer which is consistently ranked the world's second largest appliance maker by units sold. Electrolux products sell under a variety of brand names (including Frigidaire and Electrolux) and are primarily major appliances intended for consumer and professional use.


Extract Document - AI

Goals

  • Extract Data - AI
    To build a solution that maximizes sales revenue across distribution channels by optimizing marketing and promotional spend.

Extract PDF - AI

Approach

  • Extract Document - AI
    Used historical sales and marketing data to build a multivariate linear regression model.
  • Extract File - AI
    Applied logarithmic, linear, power, exponential, box-cox, and gamma transformations to improve data distribution.

Extract Document Data - AI

Results

  • Extract PDF - Document
    Direct increase of $1m in net profit over a 2-year period.

Background

As the world’s second largest appliance maker, Electrolux spends a great deal on through its marketing and promotional channels. Electrolux must balance its spending across TV & newspaper advertisements, billboards, promotional offers, sales representatives, and flexible pricing. Before engaging with Foundation AI, this process was conducted manually.

Challenge

Electrolux’s marketing team contacted Foundation AI to identify the optimal amount of money to spend across each of their promotional channels. Foundation AI configured its Extract Learning Platform to customize a solution to help them understand the financial impact of investing in each promotional channel.

Electrolux sought a solution that would automatically redistribute the current budget across promotional channels to optimize sales with the same total budget allocation.

The solution included:

  • Document - Data

    Scenario Analysis – A tool that helps Electrolux model how sales will be affected if they change how much they invest in each promotional channel.

  • Extract Document - artificial intelligence

    Contribution Plots – A descriptive dashboard in which percentage contribution from each channel isplotted on a pie chart. This helps Electrolux understand which channels are driving sales and which channels can be defunded without affecting sales volume.

  • Extract Data - AI

    ROI and MROI plots for each channel – Two descriptive dashboards that help Electrolux visualize ROI and MROI across channels. ROI plots help the marketing team understand the efficiency of each channel. This enables Electrolux to maximize dollar sales revenue with the same marketing spend. MROI represents the ratio of incremental sales to incremental budget allocation. The better the MROI, the more sales will increase with the next dollar invested.

  • Document - artificial intelligence

    Dollar Sales Calculator Slider Tool – A tool that lets the user adjust the marketing budget spend across various channels and automatically models how sales will be affected.

Solution

Models

Market Mix Modelling uses multivariate linear regression on historical sales and marketing data to estimate how changes in spending across promotional channels will affect overall sales. We modeled Product Sales / Product Share for different marketing channel activities including TV & newspaper advertisements, billboards, promotional offers, sales representatives, and flexible pricing. These activities were considered independent variables in the model. To increase prediction accuracy, we applied appropriate log / power / exponential transformations on the independent variables based on their dependency with Product Sales / Product Share.

Input Diagnostics

Before running the regression model, we needed to first make sure that the input parameters were truly independent and would not bias the output.

To do this we used the following:

  • Document - AI - Data

    Density distributions for sales and promotional channels helped us understand outlier distribution and which filters/transformations to apply to improve variable distribution.

  • Extract AI - Data

    Correlation plots between sales and the promotional channels helped us understand how they are correlated and helped us identify if we needed to transform a variable before feeding it to the model.

  • Data Extraction

    Correlation matrices quantified the correlation between y and x variables and helped us identify interdependencies present between independent variables. We then used appropriate transformations to eliminate these interdependencies.

  • Data Extraction - AI

    Time series plots for sales and promotional channel activities helped us understand seasonality and moving trend factors for each variable.

Regression Model

Using density distributions, we removed outlier data that might skew the results. From the correlation plots, we checked if sales data was related to promotional channel data and used either logarithmic, linear, power, exponential, box-cox, or gamma transformations to improve the distribution.

We then ran a multivariate linear regression in which sales was regressed on promotional channel data. Based on the P-values and R-squared of the model, we reconsidered which variables to include in the analysis and what transformations to use on the variables.

We then validated the regression using the following methods:

  • Document Data Extraction

    R-Squared - R-Squared of a model determines how well our model explains the variations in dependent variable by changing the independent variable—the better the R-squared, the better our model.

  • AI - Data Extraction

    P-Values - P-Values of an individual independent variable (X) in a model represents probability with which we can disregard what is called the null hypothesis. The null hypothesis in regression means that the dependent variable (Y) has no dependence on a particular independent variable (X). A smaller the P-Value indicates a better model. The ideal P-Value for a variable should be <=5%.

  • Extract

    Error metrics – The quality of any predictive model is determined by the degree to which its predicted values are comparable to that of actual values. This can be quantified using functions like Correlation, Root Mean Squared Difference, Mean Absolute Percentage Error, Mean Squared Error between predicted and actual variables.

  • artificial intelligence - Document data

    K-Fold Cross Validation - This is done to ensure that our model is not over-fit and that it would produce similar results with new data points. To conduct K-Fold Cross Validation, our data was split into “k” parts, regression was performed on “k-1” parts, and the results were tested on the “k”th part. This was done “k” times by changing the test data set every time. In every split regression, the R-Squared, P-Values, RMSE, MAPE and MSE were similar to the results of the regression model that was run on all the data.

Results

Implementation of this tool has led to a direct increase of $1m in net profit over a 2-year period.

If you are interested in deploying a solution built on our Extract Learning Platform, contact usto see what Foundation AI can do for you.
Artificial Intelligence for the Real World
© 2021 Foundation AI