RESTBank is launching a new product - Home Insurance. Management Team already decided about the most of the Advertising Campaign details but they are not sure which of the 8 Ad Projects will be most successful in convincing customers to buy an Insurance. Ad Campaign is already low on budget, so long trials are not an option. Solution has to be swift, with proven effectiveness and low-cost. AI Team came up with an idea of building an algorithm that will just do the thing. Implementation will cost nearly nothing and will enable to determine what is the best advert within couple of hours. Challenge accepted?

The Trial

The trial will last 10.000 rounds. Each round is a customer that enters RESTBank website. Next we display to that person one of the designed Ad Projects. Customer either click it (success) or not (fail). Project with most clicks win. We'll compare two methods - first when we display Random Advert and the second where the A.I. Algorithm controls ad selection.

The Environment

Let's pretend that we have a Crystal Ball for a moment. Below, you'll define Customer Ad Preferences. So we'll know from the beginning which ad has the best conversion rate. But we won't pass that knowledge to the A.I. Algorithm. If it can figure it out by itself. It has to be significantly faster than Random Method if we want to convince the Management.

The Success Criteria

We'll compare absolute and relative returns between Random Sampling and Thompson Sampling methods. RESTBank's profit on each insurance is 1000 USD and that's how much is worthy one customer difference between two methods. So if the A.I. Algorithm can figure out the most effective Ad after 2.000 customer visits, it'll drastically decrease campaign costs.

Predefine Statistical Customer Preferences

{{ error }}
##### Peek into Crystal Ball - proposed advert designs with conversion rates

So what are the Ad tastes of an statistical RestBank Customer? According to previous page survey we can clearly see now which Ad will be able to convince 21% to buy a Home Insurance. But in the real world we won't have a crystal ball - so it's just an assumption for the simulation purposes. That's why AI Algorithm won't have that knowledge. It'll have to figure it out through observation of customer's consecutive choices.

10.000 Rounds

With every round Thompson Sampling algorithm will narrow the choice of displayed Ads. In simplified example, if among first 100 customers none of them clicked on project "Family + Car", then that proposal will be greatly omitted during the rounds 101-200 in a favor of more popular chioices.

The Benchmark

Random Sampling - that's what we play against. On each of 10.000 rounds we'll display random Project to our customer. Bottom line is why do we bother our Management with this whole "AI" if we can't do significantly better than Random...

What A.I. doesn't know?

We'll compare absolute and relative returns between Random Sampling and Thompson Sampling methods. RESTBank's profit on each insurance is 10.000 USD and that's how much is worthy one customer conversion.

Conversion Rates according to customer preferences

Conversion Rate: {{ scores['female_trust_cash']}}%
Girl + Trust + Cash Bonus

Conversion Rate: {{ scores['female_trust_car']}}%
Girl + Trust + Car Insurance

Conversion Rate: {{ scores['female_familyhome_cash']}}%
Girl + Family Home + Cash Bonus

Conversion Rate: {{ scores['female_familyhome_car']}}%
Girl + Family Home + Car Insurance

Conversion Rate: {{ scores['family_trust_cash']}}%
Family + Trust + Cash Bonus

Conversion Rate: {{ scores['family_trust_car']}}%
Family + Trust + Car Insurance

Conversion Rate: {{ scores['family_familyhome_cash']}}%
Family + Family Home + Cash Bonus

Conversion Rate: {{ scores['family_familyhome_car']}}%
Family + Family Home + Car Insurance

##### Thompson Sampling Algorithm vs. Random Choice

As we see on graphs below, our AI algorithm was able to figure it out much faster and with greater precision that Ad Project "{{thompson_list[0][0]}}" will have the greatest conversion rate. Algorithm decided to display it {{thompson_list[0][3]}} times to customers (compared with {{random_list[0][3] }} of Random Sampling), achieving overwhelming confidence in superiority of Project "{{thompson_list[0][0]}}". Just with the small sample of 10.000 bank customers we're able to save {{relative_return}}% costs compared to Random Sampling benchmark method. Enough to impress Management?

Conversion Rate: 21%
{{thompson_list[0][0]}}

The Time

Thanks to Probability Matching Strategy AI was analyzing feedback received from the Customers in a real-time. It didn't had to wait until the very last round to draw conclusion.

The Profit

Just with the small sample of 10.000 bank customers we're able to create {{absolute_return}} USD absolute return which was {{relative_return}}% bigger than return achieved with Random Sampling.

AI. Algorithm vs. Benchmark Comparison