In the last instalment of our series exploring the inner workings of Ibex, we will be focusing on the final three components of the decision-making process; the Lifetime Value (LTV) model, Experimentation module and Control group module.

At this stage of the decision making process, Ibex can now accurately model the casino income generated for each combination of players and marketing actions.

The final steps involve mapping short-term behaviour onto long-term behaviour, experimentation and the allocation of control groups.

LTV Model

Ibex predicts the lifetime value of players for every possible action, which is the future value we think the player will generate after an action is sent.

More than 95% of player value is generated within 180 days of sending an action, therefore when we discuss lifetime value we are referring to this 180-day period as this simplifies the analysis and allows us to more quickly gather our training data.

Ibex models lifetime value by considering short-term (ST) and long-term (LT) effects in isolation. This allows us to monitor the outcome of sending an action for just a short period of time.

In the short-term (i.e. day 0 and day 1), the focus is on whether the player responds to an action and the associated revenue and costs as a result.

In the long-term, we are interested in any value generated after this short-term period (i.e. day 2 to day 180). Although the value generated in this long-term period will not directly stem from the action sent, it is closely related to the success of the action because retaining a player increases their expected future value.

The following chart emphasises how a positive short term outcome increases long-term value dramatically. D = 0 is the long term value given a player does not deposit in the short term and D = 1 is the long term value given a player does deposit in the short term.

By summing the short and long-term value we calculate lifetime value:

The short term value prediction is the focus of the Deposit Rate, Deposit Amount, GGR and Cost Models explained in the previous instalments of this blog series.

The long-term value prediction, is modelled by three key predictors:

1. Whether a player deposits in the short term
2. Recency (how long since the player was active before the event occurred) – when a player is less active, this decreases their lifetime value
3. Historic value (how valuable the player has been historically) – when a player has displayed high value historically this increases their lifetime value

Using these three key predictors we optimise for the impact in Lifetime Value by taking the difference between Expected LTV given a player deposits (D=1) & the expected LTV given a player doesn’t deposit (D=0). This is then multiplied by the difference in the deposit rate given an action is taken (T=1) & the deposit rate given no action is taken (T=0).

Therefore, if taking an action increases the short-term deposit rate, then there is a positive long-term value impact.

The main reason we have modelled LTV in our decision making process is because in some cases Ibex will see value in taking short-term risks. In other words, a positive LTV impact means that in some situations it pays to accept a short-term loss. This highlights the importance of retaining players and quantifies how much we should spend in doing so.

Experimentation Module

The experimentation module ensures that any new actions added to Ibex are working to their best of their potential. In order to achieve this, this involves experimenting with a small number of offers each day.

3 stages of action experimentation

There are three stages of experimentation in the lifecycle of an action.

1. Optimised – We have full confidence in our deposit rate model for this action. No experimentation will be performed
2. Hybrid – We do not have full confidence in our prediction for this action. However we have full confidence in an associated action that is the same action type
3. Experimentation – We do not have a stable model for either the action or the action type. We will fully experiment with the action.

In Ibex, the term ‘action type’ means similar types of offers i.e. match bonus offers, deposit spins offers, cashback offers etc.

The following diagram illustrates how an action is assigned to one of these 3 stages.

Selecting customers to experiment with

Ibex only experiments with customers who are eligible to be contacted on a given day and these customers are only selected after we have already determined the next best action to take for each of them.  

A customer will either receive the next best action (which could be doing nothing) or an action from the randomisation pool, however this is not an equally weighted pick.

If the action picked for a customer is calibrated and fully optimised (stage 1), they will receive this offer at least 50% of the time but this is dependent on how many actions are still in stages 2 and 3. Over time, this percentage increases as more actions become optimised.

Winding down experimentation

Over time, Ibex gradually reduces the experimentation it performs on actions. In the first 2 weeks, all actions will be considered for experimentation.

After this, Ibex reduces experimentation if

1. Enough recommendations have been sent out to start optimising
2. The offer has been active for 6 weeks

Once one of these two criteria are met, the experimentation is ramped down until that action is no longer experimented with.

Control Group Module

The allocation of control groups is the last step that Ibex takes in the decision making process. This is a key component as it allows us to quantify the impact Ibex has.

Therefore, once Ibex has selected the best action for a player, there’s still a small chance (~20%) that the player will not actually receive an offer / message at this point in time. Players in the control group are players that

• Would have received an offer based on the Ibex algorithm
• Have been randomly selected to be put in the control group

Its important to note that players are only in the control group for one specific action, i.e. the same player that is placed in a control group today might still receive an offer in a few days.

This results in a steady stream of players in the control group over time.

We can use comparisons between campaign / control groups to quantify the impact has on a true ‘do nothing’ group on a number of different metrics such as deposit rates, deposit amounts, NGR, Income etc.

Now these final three steps have been taken in the decision making process, Ibex will choose the next best action for every contactable player each day.

So, to summarise all the steps of the decision making process:

1. Determine all players which can be contacted with marketing actions
2. Predict the deposit rate impact
3. Predict the deposit amount uplift
4. Determine the GGR generated
5. Determine all costs associated with sending marketing actions
6. Predict lifetime value impact
7. Experiment (if necessary)
8. Allocate a small % of players to control groups

Ibex is going through each of these steps for every possible combination of contactable players and marketing actions each day.

In conclusion, Ibex provides a powerful tool for iGaming platforms looking to harness AI to make optimal marketing decision for their players. By following the decision-making process outlined above, Ibex can accurately predict the impact of marketing actions on short-term and long-term value and help iGaming companies make informed decisions on where to invest their marketing resources.