Pricing decisions are crucial for any business, as pricing is directly linked to consumer demand and company profits. Consequently, sales forecasting is critical for every company, especially in the food industry. The objective of this project was to predict sales and estimate the impact of promotional initiatives in an effort to optimize pricing. Sotiris Archontas created and tested sales forecasting models and concluded that the impact of own pricing was far greater compared to the impact of competitor pricing. Furthermore, there is a delay between the promo event and the increase in sales as well as a saturation point, after which promotional events no longer impact sales. Model testing revealed that sales forecasting can be significantly improved using larger and more detailed data sets.

Case Study Contents

1. Background
2. The challenge: Optimizing sales for Food Manufacturer
3. Getting started
4. The workflow

 Creating prediction models
 Data used for Preprocessing

5. Phase 1: Studying the impact of selected parameters on sales

 Impact of product price on sales
 Impact of competitors pricing on sales
 Impact of ad promotions on sales

6. Phase 2: Creating Prediction Models 

 Model 1: Sales as Time Series

7. Phase 3: Testing our Sales Prediction Models

 Testing Model 1
 Testing Model 2
 Feature Engineering: Impact of promotion

8. Key Takeaways

9. Online presentation
 

1. Background

Witside is a technology company using Data management, Data Visualization, and Advanced Analytics to provide consultation services. Witside helps companies "measure tomorrow" and use data to achieve cost optimization, additional revenue, optimized processes, and enhanced customer experience. The company offers end-to-end forecasting solutions combining internal historical data with outside data to improve the accuracy of forecasting models. Witside's work is based on a unique framework consisting of 4 steps:

 Step 1: Validate Data

 Step 2: Produce Baseline Forecast

 Step 3: Embed exogenous variables and produce actual forecasts

 Step 4: Operationalize

Deploying AI-driven software solutions Witside enables companies to perform effortless forecasting, embedding their personal experience or calendar events that affect its accuracy.

The objective of this project was sales optimization for a Food Company using AI-driven predictive modeling.

Sales optimization in the food industry is a multifaceted topic, and Witside, in collaboration with Big Blue Data Academy, assigned Sotiris Archontas the sales forecasting and optimization project. Using the data provided, he created models that enabled the extraction of information about parameters affecting sales.

Note: For data protection reasons, the identity of products is not presented, and the sales volume is scaled or hidden.

2. The challenge: Optimizing sales for Food Manufacturer

Sales forecast is a critical business factor for companies of all sizes, as it enables them to forecast demand and develop effective pricing and promotion strategies. In addition, given the complexity of pricing in the dynamic market conditions of the food industry, modeling can provide valuable insights that help companies improve their decision-making processes.
 

3. Getting started

At the initial stage of our project, we studied the effect of the following parameters on sales:

 The prices set by the food company that assigned the project.

 The prices set by competitors in the market.

 The presence or not of advertising campaigns per week.
 

4. The workflow

Creating prediction models.

The first step was to create models that predict future sales and estimate the effect of the advertising promotions. To perform the analysis, we used data that included

 Volume of sales

 Client product prices

 Competitor product prices

We also used data relevant to the advertising promotions of the company:

 The volume of sales that took place the week after the promotional event.

 The impact of promotions on sales of company products
 

Data used for Preprocessing

The data used for preprocessing included 6 product categories including:

 28 customer products

 1374 competitors' products

The available data had the following duration:

 The duration of sales data was 156 weeks data for sales (3 years)

 Promo data had a duration of 52 weeks data (1 year)
 

Initial analysis: finding a correlation between client and competitor products.

The first challenge of the project was to establish a correlation between our client's products and the competitors' products.
 

5. Phase 1: Studying the impact of selected parameters on sales

Impact of product price on sales

We studied the effect of product price on sales over a year and had the outcome presented in Graph 1 below.

Graph 1: Impact of product price on sales
 

Darker color dots signify lower prices, and we can see in Graph 1 that low prices lead to increased sales - a finding that was to be expected.
 

Impact of competitors pricing on sales

We studied the effect of competitor prices on sales over a year and had the outcome presented in Graph 2 below.

Graph 2: Impact of competitors pricing on sales

 

Comparing our sales against competitor pricing showed no correlation between the two parameters. In theory, we would expect our sales to increase when the competitor price goes up, but we didn't see something like that for the duration of our data.
 

Impact of ad promotions on sales

In Graph 3, we see the volume of sales in correlation to the occurrence of a promo ad.

Graph 3:  Impact of ad promotions on sales
 

 Blue dots: weeks with ad promo on tv

 Red dots: weeks with more than one ongoing promo – tv, digital, etc.

 Black dots: weeks with no promo
 

Conclusions based on Graph 3

Conclusion 1 – Time-lapse

Based on the data displayed in Graph 3 there is a time-lapse between a promotional event and an increase in sales. Thus, we see sales increasing 3 to 4 weeks after the advertising promotion took place.

Conclusion 2 – Promo saturation

Based on the analysis on Graph 3, it becomes apparent that promo campaigns need to have a specific duration for optimized results. After a number of weeks, we see a saturation point where sales decrease despite promo campaigns.
 

6. Phase 2: Creating Prediction Models  

Following our analysis, we created and tested predictive models for sales forecasting.
 

Model 1: Sales as Time Series

For our first model, we approached sales as time series, as seen in Graph 4 below.

Actual vs. Predicted sales

Graph 4: Time Series Prediction

 

Blue line: Real-time series for sales for a 3-year duration.

Orange line: Prediction provided by ARIMA model 

Time Series prediction – results

Based on Graph 4, we conclude that initially, we have a relatively accurate prediction that captures the peak in actual prices. The prediction accuracy of the model is considered to be good. However, over time there is a lapse to the right of the prediction line compared to the actual. This is attributed to the fact that we don't have sufficient data that would enable us to build a more accurate Time Series Prediction Model. 
 

7. Phase 3: Testing our Sales Prediction Models
 

Testing Model 1

For Model 1, we used Random Forest Regressor.

Input data for Model 1 were:

 Prices of our client company

 The price difference between our company and the competitor

 Competitor Sales for a specific product

Result/Output: The expected/predicted sum of sales.
 

Model 1 Optimisation

In order to optimize Model 1, we excluded Trend and Seasonality from the original Sales Time Series.

Graph 5 Decomposition of seasonality and trend

By removing a small percentage of seasonality from sales, we saw an improvement in the model and concluded to R2 = 0,69
 

Testing Model 2

For our second testing model, we added to our modeling one more input feature: the existence or not of an ongoing promotional campaign.

Input data for Model 2 were:

 Prices of our client company

 The price difference between our company and the competitor

 Competitor Sales for a specific product

 AND the existence or not of an ongoing promotional campaign.

Result/Output: The expected/predicted sum of sales.

The result of our model was R2 = 0.59

This comes in correlation with the analysis of

 the effect of advertising.

 the necessary duration to have a real impact on sales.
 

This feature type is very simple, so we created more columns to help the model better understand the impact of the ad promotions.
 

Feature Engineering: Impact of advertisement promo

The new columns included:

 Sales shift

 Consecutive days of promo

 Consecutive days of promo shift

 Moving average

Conclusion: We did not observe an improvement in the model as the available data was not sufficient. Data provided on promotional events only spanned over a year – 52 weeks, so our model had only 52 rows.
 

8. Key takeaways
 

 The impact of our company's pricing was far greater compared to the impact of the competitors' prices.

 We observed a delay between the promo event and the increase in sales.

 There is a saturation point, after which promotional events no longer impact sales.
 

Recommendations to improve predictive models

 More sales and promo data would significantly improve our predictive models.

 Having the same promotional campaign happening at different time periods would provide valuable data, as it would help us understand if the increase in sales is attributed to promo or seasonality.

 Daily data would help a lot towards training our models compared to using weekly data. 
 

9. Online presentation

You can see Sotiris Archodas's final data science project presentation here.