amazon_-eCommerce_optimization

This project includes an analysis of several advertising campaign results of big consumer brands on amazon.com, with a goal to look for potential ways of optimization to improve ROIs.

Technology I used:

• Prophet time series to forecast potential sales.
• Random Forest Regressor from Sklearn to predict factors affecting sales
• SHapley Additive exPlanations (SHAP Values) algorithms for the model interpretation.

Project goals:

1. Analyze existing data to find best practices of advertising (slice by company / portfolio / ASIN).
2. Identify campaign optimization changes with a focus on ROI increase by adjusting:

• Keywords
• Bid management
• Time of day / week
• Placements
• Campaign type

3. Forecast potential sales with a budget increase.

Data

• MySQL DB in AWS
• 13 tables
• 104 campaigns
• Aug 2019 - Feb 2020

Campaigns EDA

I calculated the average results of all campaigns to have a baseline:

Impressions: 10M Clicks: 23K Spend: $32K Sales: $92K ROI: 2.875

Campaign types

Amazon has 3 types of campaigns: Sponsored Brands (SB), Sponsored Product (SP), and Sponsored Display (SD). I looked into ROI by campaign types, keeping in mind that average ROI for all campaigns is 2.875, and found Sponsored Brands campaigns have ROI 7.056!

My first recommendation: invest in Sponsored Brand campaigns.

Targeting

When using automated targeting, ROIs decrease.

My second recommendation: use manual targeting.

Ads placement

We can also optimize campaigns via adjusting ad placements.
Furthermore, ROIs vary for different ad types.

For brand ads, top search works much worse than other placement: 7.07 vs 12.32 ROIs.

For product ads, top of search ad placement on the first page of results perform best: 4.317 vs 2.083 on the rest of the search. Ads on product pages and off-amazon remarketing show poor performance: 1.094 and 0.791.

Search EDA

Let's take a look into keyword analysis. My goal was to analyze each keyword performance dynamically. For this part, I used plotly library.

Keyword performance

The charts below show the performance for a search term 'hand sanitizer' across all campaigns.

Future sales prediction

To extrapolate future sales, I used Prophet time series. Below is a graph that predicts sales coming from the keyword 'hand soap' for two upcoming months.

We can see overall trend for sales from this keyword, and weekly fluctuation component.

Product analysis

Next, let's see how each product is performing. Below is a brief overview of one product, hand sanitizer's, performance. The first image shows the funnel impressions -> clicks -> orders. The second shows overall money spent on ads -> sales.

Product performance by campaign

I performed high-level analytics of data grouped by campaign. This showed how ads are performing for product A (hand sanitizer):

Product performance in dynamics

Below are the graphs for impressions, clicks and sales for the product A, hand sanitizer, across all campaigns. There is also an option to look into each campaign individually.

Sales forecast

Using Prophet time series, I build sales forecast for each product. The graph below shows the two-week hand sanitizer sales forecast, overall trend, and weekly fluctuation component.

Factors, affecting sales

Finally, I wanted to look into factors affecting sales. To do that, I built a ML model (Random Forest Regressor from Sklearn), and used SHAP values to visualize features importance.

I used the following features to predict sales:

• impressions
• clicks
• CTR
• CpC
• money spend
• portfolio
• daily budget

Each feature has different affect on the outcome (sales). A high-level view can be drawn from the model itself, by plotting these feature importance parameters:

importances = rforest.feature_importances_
indices = np.argsort(importances)
features = Xtrain.columns
plt.title('Feature Importances')
plt.barh(range(len(indices)), importances[indices], color='b', align='center')
plt.yticks(range(len(indices)), [features[i] for i in indices])
plt.xlabel('Relative Importance')
plt.show()

This graph shows number of clicks and money spent are the most important (which makes sense), but how exactly do they affect sales? To understand the model behavior better, I looked into SHAP values.

SHAP values show the importance of each feature by comparing what a model predicts with and without the feature. In other words, each SHAP value measures how much each feature in our model contributes, either positively or negatively, to each prediction.

SHAP summary plot shows the positive and negative relationships of the predictors with the target variable (sales) for each factor. Each dot represents one observation of training data. Below is a short graph explanation.

• Feature importance: Variables are ranked in descending order.
• Impact: The horizontal location shows whether the effect of that value is associated with a higher or lower prediction.
• Original value: Color of the dot shows whether that variable is high (in red) or low (in blue) for that observation.
• Correlation: A high level of the “click” feature has a high and positive impact on the sales. The “high” comes from the red color, and the “positive” impact is shown on the X-axis. Similarly, we will say high number of “impressions” is negatively correlated with the target variable (sales).

We can look in depth into each feature's behavior. For example, let's see how the cost per click affects our sales.

This graph is called 'The partial dependence plot'. It shows the marginal effect that one or two variables have on the predicted outcome. It tells whether the relationship between the target and the variable is linear, monotonic, or more complex. In our case, the change of cost clicks doesn't seem to affect the outcome (sales) much, though there is a slight linear dependence in between them. SHAP also shows that CpC affects how much money we spent.

For each observation, we can look into plot for all features called force plot. Below is a plot for our 12th observation. In this case, total sales were $70.99. The model's base value (average) is $54.44, and the major factor that pushed the sales up were clicks - we had 7 of those.

Summary

The plot below shows simple interpretation of all feature mportance’s and effects. The red color means a feature is positively correlated with the target variable. In this model, high clicks, CTR, CpC and daily budget will increase sales.
High impressions lower the sales down.

Next step:

Moving on, I'd like to create an interactive webpage with the features described above to provide clear and detailed information about each advertising campaign, search term or product.