Calculating Cost Per Action for affiliates

This is the first article in the series about CPA. Let’s discuss how mathematical models can be applied to affiliate marketing in trading and what risks should be taken into account.

What is affiliate marketing?

Affiliate marketing is a type of marketing in which a company pays affiliates a commission for advertising its products. The results are tracked via a unique affiliate link that tracks traffic from the affiliate’s website or blog to the company’s website.

Affiliate marketing is a great way to attract new clients and drive sales. And its popularity is growing. According to statistics, the affiliate marketing industry is set to grow to approximately $13 billion in 2022 and $15.7 billion by 2024.

It’s not only e-commerce giants, like Amazon or eBay, that benefit from affiliate marketing. Exness and many other brokers have also developed their own affiliate programs, and most of them use the CPA model to calculate an affiliate’s commission.

CPA, or Cost Per Action, means that an affiliate gets a fixed reward after the client performs a specific action, e.g. deposit, trade.

In this article, we will discuss how mathematical models can significantly improve your CPA affiliate program and which risks should be taken into account. The learnings below are mostly applicable to the trading environment. However, they may ignite inspiration for those in other industries as well.

Mathematical model implementation

If you want to make your program more attractive to affiliates, consider these three factors:

1. Reward value
2. Payout speed
3. Reward transparency

Reward value

Clients attracted by an affiliate can differ a lot in terms of value. Logically, it’s better to pay bigger rewards for the more valuable clients, rather than paying a blanket commission for all.

However, how can we tell a promising client from one that won’t bring us any value? Usually, we look at the deposit amount: the bigger the first deposit, the higher the chance that the client is truly interested in trading. Thus, the commission paid for such a client should grow proportionally. This sounds fair and attractive to the affiliates.

Unfortunately, it sounds attractive to fraudsters as well. The CPA model implies the “paying in advance” principle, so there’s a high risk of clients withdrawing their money soon after affiliates get a reward.

Did you know? More than 67% of brands worry about affiliate fraud and close to one-third have been affected by it.

Affiliates revenue model

To deal with this risk we can apply the following logic (pic. 1):

1. Collect cumulative revenue for several days.
2. Based on the speed of the revenue increase, approximate the future revenue (e.g. using trend continuation or spline).
3. Pay rewards to affiliates if the client’s future revenue exceeds a certain minimum amount.

pic. 1 (a) — Payment curve modeling example without payouts

pic. 1 (b) — Payment curve modeling example with payouts

Let’s consider the following example (pic.1):

• For each client brought in, the reward is $18.
• An affiliate brought in 1 client and logically expects to get $18.
• We pay this reward if the modeled future value for the client is greater than 15$.

Based on the client’s activity for 1–3 days (pic. 1a) the modeled curve grows slowly and shows revenue estimation of only $15.

However, if we wait for 2 more days (pic. 1b), we see the growth increase. The future value is now $20 and we can pay a reward for one client ($18) after 5 days of delay.

This approach may limit fraud cases, but it has one significant flaw. As mentioned earlier, payout speed is hugely important for affiliates. If it depends on the client’s activity and the revenue curve grows slowly, affiliates won’t know when they get a reward and why it is delayed. So this is not a great solution and it can’t be used as CPA.

Let’s see if we can solve this problem.

Payout speed

If we want to pay fair rewards to affiliates faster, we should secure ourselves with a combination of robust anti-fraud techniques. Applying a smart data-based approach may be of use here.

We suggest using two methods:

1. Parametric curve. A curve with flexible parameters that allow us to increase the payment curve growth, based on the affiliate reputation and historical performance.
2. Predictive model for attracted clients. A machine learning model that can predict clients’ lifetime revenue right after deposits. Instead of approximated revenue, we can use the sum of predicted revenues.

Let’s look at both methods in detail.

Parametric curves

Among the flexible curves with simple parameters the most convenient for this purpose is the Gompertz curve.

where

• a — is an asymptote (pic.2)
• b — shifts graph by the x-axis (pic.3)
• c — growth rate (pic.4)

pic.2 — Varying a

pic.3 — Varying b

pic.4 — Varying c

For affiliates showing quality traffic we can easily set the revenue curve (pic. 5):

• Increasing the revenue curve growth
• Increasing the approximated future revenue

pic. 5 — Payments, using Gompertz curve

In this case (pic. 5) we set the same parameters as in the first example above. But if we see rapidly growing revenue, we can set the payout function with very fast growth. In this example, the affiliate will be paid $15 after three days, which is almost twice as fast.

However, there are drawbacks to this approach:

• It doesn’t speed up payouts for new affiliates.
• It’s hard to understand the logic behind it, so we get a lack of transparency.
• We can’t totally eliminate fraud.

Predictive models

Lifetime value prediction is one of the classic machine learning problems since every business wants to have a clear understanding of

• How much revenue will a customer generate?
• Are they going to be loyal?
• Are they fraudulent?

The developed model should be able to predict clients’ future revenue considering just their first actions. Having the sum of predictions we can dramatically speed up payouts.

However, there are still some risks associated with this approach:

• Prediction errors come at a cost.
• The same problem with transparency as for the parametric curves. It’s hard to explain the logic to the affiliates.

As far as we can see there is a trade between payout speed and transparency.

Transparency

Now we come to the last, but not least, variable in a successful CPA program, which is transparency. Our affiliates should clearly understand the rules of the game and how they can earn more.

Although 100% of transparency is hard to achieve, here are some ideas on how we can increase it:

• Implement clear basic rules for rewards amounts.
• Provide affiliates with a clear explanation of any payout delays.

Of course, we can’t reveal our anti-fraud algorithms to customers. But we can create a visual representation of the traffic quality and make it accessible for affiliates only.

Conclusion

If you want to make an attractive affiliate program, you should consider these three factors:

1. Reward value
2. Payout speed
3. Reward program transparency

And, of course, you should secure the company from fraud.

These mathematical models can help you pay fair rewards, and improve your payout speed and security. Yet they make things non-transparent, so it’s necessary to provide affiliates with additional information about the process.

In our next article, we will share our experience on how we used a predictive model to increase the payout speed.