The goal of this tutorial is to give a quick overview of the features contained in the Akur8 insurance modelling software.

This tutorial is split into two main parts.

In the first part, we will build a full model from scratch on the demo dataset. This modeling will include the elements of a state of the art insurance model such as:

In the second part, we will show some Advanced features that are targeted to more complex use cases, or show how to automate some manual processes in order to increase productivity and reduce friction.

Everywhere on the Akur8 platform, you will notice a chat icon located in the bottom right. This chat allows the user to contact Akur8’s team of Actuaries and Data Scientists that are available to answer any doubt or question you have about the software.

Do not hesitate to contact us if you are unable to proceed inside this tutorial: the team will be happy to quickly answer your questions.

To access Akur8, enter https://beta.akur8-tech.com/ in your Chrome browser.

Login using your akur8 email and okta.

After logging in, you access the workspace selection page. This page allows you to select in which workspace you will work. Workspaces are dedicated to teams working on similar models, sharing their Databases and modeling projects. If you have only access to one workspace, you will directly enter it without going through this stage.

For the purpose of this tutorial, you can enter a workspace previously created for you under your name : FIRST_NAME LAST_NAME.

The Akur8 software currently has 5 sections:

Data: to upload Databases, inspect them, and prepare the data before starting the creation of models.

Risk: to build models once the Data is ready (cf. Risk Section).

Demand: to aggregate frequency, severity and propensity models into a single risk estimate and run loss-ratio studies.

The pictograms on the left of the screen indicate which section the user is currently in. After entering the project, the user is in the Data section.

The Data section contains a list with all the Databases previously loaded in Akur8.

When entering your personal workspace for the first time, this list can be empty.

A new Database can be created by clicking on the upload database button in the middle of the screen (or at the + button at the bottom of the databases list):

Enter the name of a new Database, select a file to upload and validate the checkbox. The data can be uploaded (middle) by clicking on the UPLOAD button. The upload progress is displayed on the screen (right).

A database can be uploaded with several formats: CSV, parquet, zipped CSV (for fast upload of large databases) or SAS database.

Once the Database is uploaded, it must be pre-processed, to compute distributions and statistics. This pre-processing may take a few minutes.

The progress bar showing the pre-processing status is displayed at the top-right of the screen.

As the pre-processing is running on the cloud, it is possible to upload other datasets while the process is running.

Once a Database has been uploaded and pre-processed, its summary is displayed.

This summary gives the main information on the Database: name of the original CSV file, creation date, and some statistics on the number of rows and columns.

It is possible to share information about the database in the Documentation note. This information will be available to all users granted access to the workspace, and will be included in the final documentation of the models (see section Exporting your models & documentation). It is also possible to see who was in charge of the creation and last modification of the datasource.

The list on the left contains all the variables present in the database. We can select each variable for a detailed inspection.

It is possible to search for a particular variable by clicking on the search icon on the top of the list. For example, let’s type age on the search box to select the driver_age variable.

The top of the screen represents the distribution of the selected variables. Missing values are highlighted in red. In the center, you can find some statistics describing the variable (some statistics such as the median and the mean are available only for numerical variables). A documentation, giving information for the selected variable, is also available.

The variable type is defined at the right of the screen:

Ordinal variable: if the variable follows a specific order. All the numeric variables are considered ordered by default, but sometimes non-numeric variables should be set as ordered as well (for instance a variable containing levels “A”,”B”,”C”… might have a relevant order).

Categorical variable: if the variable contains categories that have no specific order. All the non-numeric variables are considered categorical by default, but sometimes numeric variables can be categorical (for instance if you have 3 groups “1”, “2” and “3”

with no specific order).

The type of variables (ordinal / categorical) will have a strong impact on the models created: if the variables are ordered, the models created will be smooth along it. Otherwise each level will be treated independently.

Missing values are handled by the Akur8 solution: the corresponding coefficient will be set to the average value by default, and follow the observed data if they are significant (it won’t be grouped with its neighbours).

This process is key for the performance of the modeling, but it can be time consuming. The good news is that this process can be reapplied automatically across different data sources, see Advanced Usage - Reuse Format.

There is a keyboard shortcut to quickly review all the variables and change their types when needed: Alt + ↑ or Alt + ↓ to select the previous or next variable, or Alt + O and Alt + C to change a variable type from categorical to Ordinal and vice-versa.

A summary of these shortcuts can be found in the help section.

It is possible to modify the data of each variable (grouping, reordering and management of missing values) by clicking on MODIFY VARIABLE-> MODIFY LEVELS.

We cover in detail the data transformation part in the Advanced Usage section.

Once you have uploaded, checked and modified your data, click on the Risk section on the left.

The list on the left of the screen contains all the modelling projects already created (when entering for the first time, this list should be empty.).

In this section, our goal will be to model a Frequency MTPL Property Damage cover. The modelling will consist of different steps that will allow us to see most of Akur8’s modelling process.

You can create a new model by clicking on the Create Project button in the middle of the screen (or the + button at the bottom of the list if other projects have already been created).

Once the project has been created a summary of the modeling Project is shown on the screen with some of its features: name, type, creation date and database on which the modeling will be performed.

In order to create models, the project needs to be set-up. You must define:

The first screen allows to set the goals of the modeling project:

This step allows the user to define which variables cannot be used in the modeling. Akur8 will then detect and select the optimal subset of variables depending on the constraints defined by the user.

There are different reasons to omit variables:

A-posteriori variables , not available when the underwriting takes place: the target and exposure variables are obvious examples of a-posteriori variables, but some other variables, modified during the claims management process, may enter this category.

To ease the detection of those a-posteriori variables, which are often very strongly correlated with the target, we show a Variable Importance score, which indicates the individual predictive power of each variable.

For instance, the cost of a Property Damage claim (zero if the client has no claims) is naturally strongly predictive of the actual number of claims: it is of course an a-posteriori variable.

Unreliable variables , for which quality is too low or whose origin is unknown.

Unavailable variables that can’t be accessed in the production environment (and while they could be fetched, this would require significant changes in the underwriting process and cannot be done for practical reasons).

Variables that shouldn’t be used in the modeling can be removed with the blue switch on the left of their name. Typically the variables of Target, Exposure and Date are unselected by default.

When this has been done in one model, it is possible to reuse the variable definitions: this process is described in the reuse variable definition annex of this tutorial.

Other modeling options, impacting the variables behaviours, can be set at this stage. These more advanced options (including variables offsetting and constraints) are described in the Force a variable behaviour annex.

Now we are ready to build the models.

This step allows us to build a batch of models (GridSearch).

You must define:

The numbers chosen for parsimony and smoothness define the number of models created: larger values will mean more precise choices in the models’ creation, but also longer computation time. After selecting the variables range, you can leave the number of steps at their default values.

For each combination of smoothness and parsimony created, 5 different models are created:

The use of cross-validation allows to have a reliable estimate of the performance of the models created.

(Later on, the scores (Gini, Deviance…) displayed for the models created will be the average scores on the 4 folds. Error bars will indicate the performance of the least performing fold and the performance of the most performing fold (if you are not familiar with k-fold cross-validation, you can find information on Wikipedia).

Once the process is finished, we can see the models created in a graph, where each point indicates a different model. Since we requested 5 levels of parsimony and 7 levels of smoothness, 35 different solutions to our modeling problem have been built, with different numbers of variables and smoothness.

In the graph:

This graph allows to have a clear visualisation of the possible trade-off between model complexities and performance:

After selecting a model by clicking on it, the Model Overview screen opens.

The goal of the model overview is to give more insights about the performance of a model (on the cross validation), together with a quick overview of its properties (variables selected along with their importance).

On the left, the importance of the selected variables.

The length of each bar indicates the importance of each variable, through the spread of its coefficients.

A definition of the variables spread is provided in the Variables Spread appendix.

On the right-hand side, an in-depth inspection of the metrics is possible by navigating three tabs:

A description of some of these KPIs is given in the Technical Appendix of this tutorial.

All variables are displayed in the list on the left; they are ordered by importance in the model. The variables selected are identified with a blue point next to their name ().

A graph for the variable selected is displayed on the right. This graph contains three lines and a histogram:

Each line can be shown / hidden in the graph by clicking on their name in the legend.

It is possible to zoom in by selecting the region to zoom in, or by selecting the interval, moving the mouse vertically or horizontally.

To zoom out, double-click on the graph.

When choosing a model we can find ourselves in different situations:

Akur8 allows to navigate all the different models, to select the best trade-off between these two extreme scenarios.

When reviewing the model’s variable, as soon as coefficients suspected of over-fitting are noticed, a smoother model should be selected.

For instance, in the tutorial Database, the variable vehicle_max_speed may show a tendency to over fit (its coefficient capture dubious signal): if this is the case, click on the / button to select a more robust model.

Note that the smoothing applied is consistent along all the variables in a model: clicking on the smoothing buttons will select another model, with a similar number of variables and a weaker/stronger smoothing.

High smoothness: in the previous figure, the Akur8 engine segmented the variable in just three segments. While this choice is robust, some relevant signal is not captured on the young drivers segment.

Low Smoothness: in the previous figure, the smoothing applied is insufficient as the modeled trends are not consistent. The model is clearly overfitting as it is capturing noise.

Optimal Smoothness: the previous figure shows an optimal tradeoff between robustness of segmentation and signal included in the model.

Within the Model Tree on the left of your screen, the model you just tagged will appear under the name of your modeling project.

The addition of interactions has two main parts:

The interactions considered can be chosen in two different ways:

Once the computations are done, Akur8 will show the list of interactions sorted by their statistical significance (a score below 100% means the interaction’s significance is below the significance threshold used for the GLM models creation).

In the example of the image above, the interaction between driver_gender and driver_age is the most significant. The graph shows the value of the observed and predicted as a function of the age, split into male and female gender. By inspecting the young segment, (the leftmost levels of each distribution) we can see an interaction. Young females’ risk is overestimated (orange Fitted line is significantly above the Observed line in purple), while the male one is underestimated (opposite position of orange and purple line). The Akur8 engine properly detected this (well known) effect, and associated a high interaction score.

At this stage, the user should inspect the interactions proposed by the tool, and unselect those that, even if they appear to be significant, do not have value in an actuarial sense.

As before, Akur8 displays a grid search graph showing the performance of the new interaction models. The initial GLM model (in orange) is also shown for comparison purposes.

There are three possible visualisations:

In order to create a geographic model, you will need to use a geographic database containing at least 3 columns:

NOTE: The zip code is largely the most common geographic variable encountered in data sets. However, any variable on which it is possible to attach a latitude/longitude can be used as easily. This means that if the exact location of the client is available, the Akur8 solution is able to compute Geography even with millions of distinct geographic points.

If only projected coordinates are available, it is still possible to apply geography. However the visualisation capabilities will be limited as it will not be possible to place the locations on a map.

The Akur8 engine admits the presence of zip codes without a location, up to a quality threshold (to avoid building meaningless models, no more than 10% of the zip codes in the database should be missing).

In the data section, we have already uploaded this geographic database. If you haven’t done it yet, please follow the instructions provided in the Creating a new Database section.

In the Geographic Model Preparation screen, select the columns of the zip location file.

Be sure to check Coordinate Type to be equal to Lat/Long.

Finally, if the geography coefficients should be grouped in a predefined number of levels, it is possible to enable the Quantify geography coefficients and specify the number of levels.

The number of steps defines the number of geographic models generated, all with different levels of smoothness for their coefficients. To have a larger choice of geographic smoothness, you can choose 15 different smoothness steps.

The results are presented in the same manner as for the GLM grid search.

You can see on the grid search representation two colors: the blue color represents the geographic grid search and the orange color, with a name pointing at it, is the GLM model that is being enriched.

You can see the value of the coefficient by hovering the mouse on top of them. The map is zoomable with your mouse wheel in order to better inspect the results by zone:

We can navigate through different values of the (geographic) smoothness to be able to choose the optimal tradeoff between low and high smoothness.

When we created the first GLM model (in the section Build Base GLM), we decided to exclude all regional and external variables from those considered for the model creation. This was done in order to decorrelate the pure regional effect (determined just by the position of the zip) and the external variables.

After capturing the geographic signal, we can add external or regional variables, being sure that the geographic effect determined by the policy's location is now included in the model.

The select variable page will open, as in the first creation of the GLM. By default, all the variables that are in the geo model are selected. This allows to refit the coefficients of those variables during the next fitting step if necessary. If you want to fully offset the existing coefficient and fit the external variables on top of the offset model, you can unselect all of those variables by clicking the SELECT menu icon, and click Unselect All variables.

For the purpose of this tutorial, we will keep the variables selected by default. On top of this, we can now add all the regional variables:

The popup that will appear will slightly differ from the one we previously saw on the first step of the modeling creation.

For the purpose of this demo, we will also try to include a smaller number of new variables:

After the computations, the grid search will show these results:

We can see that, in this demo database, few external variables bring value, and all the others are redundant.

Usually regional variables are known at the same level of granularity as the geographic variable (zip). For example, the regional_number_hospitals indicates the number of hospitals in a given zip code. We could take advantage of the GAM structure of the models to reduce the number of variables by including the additional information brought by the regional variables back onto the zip variable.

An inclusion score will be computed for all the variables in the model. This score indicates the suitability to reduce that variable onto the zip variable. For example the variable regional_rural_area has a score of 100%. This means that for a given zip code the variable regional_rural_area always has the same level in the database.

The new grid search with the reduced model appears. If all the reduced variables have an inclusion score of 100%, the Reduced model (in blue) will have the same performance but fewer variables than the previous model (in orange).

Classification can be performed for Vehicle models as well.

Just like before with the Geo Classification, an inclusion score will be computed for all the variables in the model so as to indicate the suitability to reduce that variable onto the vehicle_model variable.

Once you have selected the vehicle variables, Akur8 will reduce those variables onto the variable vehicle_model:

The new grid search with the new reduced model appears. If all the reduced variables have an inclusion score of 100%, the Reduced model will have the same performance but fewer variables than the previous model.

The Akur8 engine automatic grouping is based on the significance of the signal found on the data. This is a robust data-oriented best practice. However, it may happen that this is not desirable for certain segments of variables that have low exposure but a known effect.

This is for instance the case of drivers with more than 1 accident last year, as shown in the nb_claims_last_year variable. The interpretable nature of Akur8 models allows to manually adjust the impact of coefficients via the EDIT COEFFICIENTS button.

The Edit coefficients window will appear.

edit coefficients

We can move the coefficient by using the graphical interface, or by manually setting the values by clicking on the Values tab. For the purpose of this demo, we will use the “magic wand” button that will match Observed and Fitted (the last button in the Align block).

We can also modify other variables if needed by selecting them on the left side of the screen.

Three options are available:

WARNING: the goal of the Edit coefficients screen is to manually adjust segments with very low exposure, but with a behaviour known to the expert user. It is by no means meant to regroup noisy segments that may appear in the models. The data-driven approach indicates that the best practice is to first change to a smoother model until the noisy behaviours stop appearing (by using the / button on the top-right of the screen).

For the purpose of this tutorial, we selected APPLY MODIFICATIONS: as the exposure of the modified coefficients is extremely low, the change will have no impact on other variables. After the computations are done, the changes will be applied and will be highlighted throughout the modeling process, allowing complete governance and traceability.

This model will be the model that we choose as the final result of our modeling exercise. In the next step you will export its documentation.

You may want to export the models created or the associated documentation. First, the model needs to have a name (this can be done by clicking on the TAG MODEL button at the top right of the screen).

It is possible to export a model in different formats:

Many formats are available:

The documentation in PowerPoint file contains:

We have seen an in-depth methodology to build state of the art modeling using the Akur8 solution.

Akur8 provides, on top of this, some additional functionalities that can help you to do even more productive and rigorous modeling. In this section we will focus on some practical use cases:

→ Data section:

→ Modeling section:

As we have seen in the Data section, after uploading a database, it is necessary to look at the different variables in the dataset and to determine whether they should be treated as Categorical or Ordinal. This is a necessary step on the modeling that will greatly improve performance if done correctly.

If we upload a new version of the same database, or a filtered copy, it would be necessary to perform the same repetitive work twice. According to Akur8 philosophy, this task should be automated. The software allows to define relationships between columns of the dataset. This reduces the number of manual iterations required to obtain the best setup for modeling.

To link an updated version of the database with a pre-existing one, select the new database, called Tuto_DB_updated in the screen below:

A list of the columns in the updated Database will appear, together with the matched columns of the reference database, if an equivalent column has been found. A score assessing the quality of this match is also displayed.

As we can see, it is possible to rapidly spot the differences between the databases: a new column was introduced (New External Variable, at the first row), and the Matching score in the variable year is highlighted in orange, meaning that a new level (a new year) was introduced. It is possible to manually change the mapping for some columns by clicking on its name, if the column name was changed but the relationship still holds true.

All the Categorical / Ordinal properties are copied from the reference version of the database to the updated one, along with the comments and the data preparation work done on the reference database.

In the modeling section, it will also be possible to apply the same variable selection on the modeling phase between mapped databases as was applied for the initial reference database.

When creating a model, it is possible to reuse the work done in another project in the Define Variable tab.

A popup will appear allowing to choose the selected variables from other modeling projects.

Akur8 allows data processing inside the solution in order to avoid expensive iterations between the data processing software (for example, SAS) and Akur8. Akur8 by itself is not a fully featured data warehouse, meaning, for example, that it is not possible to merge different data sources.

Akur8 offers modification of column levels via grouping and reordering, and missing values management.

Consider for example the granular variable Make. It may be relevant to group such a variable by country. In the Data section, head to the View Variables page and select the vehicle_make variable.

The Modify levels screen will appear showing the distribution of the variables, together with all of its levels on the right panel. On the left there is the full list of variables. Since after a modification, the statistics are recomputed, it is strongly advised to make all the modifications to the variables before applying those modifications.

Grouping is possible by just selecting the levels to group. For instance:

The changes can be seen in real time on the graph.

Hovering over the Group Symbol of the new name group will also allow you to see all the old levels contained in it.

If you group all the makes by country, the updated variable Make will look like the picture below:

It is also possible to give a different order to the level (this is only meaningful for ordinal variables) by just dragging the level name in the appropriate position:

For instance, the studies_level variable needs to be reordered. By default, it follows the alphabetical order while the duration of the studies would be much more logical.

Finally, this window allows the user to manage missing values.

For instance, we can manage the missing values in the nb_children variable:

The missing values will be treated separately from the other levels in the machine-learning algorithm: when smoothing ordered levels, the missing values will be ignored (not smoothed) and:

In the context of a tariff update, it is often key to gradually update the tariff from a pre-existing one which is already in production. Throughout the tutorial we showed a modeling exercise completely from scratch. However Akur8 allows also to optimally update a pre-existing model. Importing an existing model is easy: right after the modeling project creation, at the Set Goals screen, at the top right of the screen, there is an IMPORT MODEL button.

Imports via Legacy Software or Json files are supported. In some EU countries, it may be important to define a Custom Format for the separator of the csv files.

It is necessary to check the mapping of the variable names in the model with those contained in the dataset. All the variables which do not show a matching score of 100% (in the example below, the variable year) must be inspected. When inspecting the variable year, we can see that the year 2020 is highlighted in red, meaning that the year 2020 is present in the Database but not specified in the model, which is logical since we are importing an old tariff without this year. By default Akur8 will try to optimally infer such value (in this case, it is grouped with 2019). The actuary has however full control over the coefficient of this new level through the edit coefficients button displayed below the variable’s distribution.

The user can access the ImportedAutoGlmKModel, the imported model. It is possible to proceed to the standard modeling and update the model.

Once we have defined the scope of the modeling project, we can apply some custom actions for a particular variable.

We can force a behaviour on this variable:

The offset can be defined graphically, by selecting coefficients and pressing the edition buttons at the right of the graph, or numerically by entering an exact value in the VALUES table.

Guaranteeing the time consistency of a model is key for correct modeling. Akur8 allows for intuitive and clear detection of significant biases in the modeling throughout the years.

Checking for time consistency of a variable is done via the interaction between the variable and, for example, the year. If this interaction is significant, the model is not consistent throughout the years.

The Akur8 solution fits interactions up to a significance value which is determined by the smoothness of the considered model. When visualizing each variable, a time consistency button is displayed below its graph. The time consistency graph shows the interaction coefficient between time and the variable, fitted on top of the pre-existing model. This allows to immediately spot time biases, avoiding noisy graphs (which are very hard to read on granular variables).

In the graph above, we can see that the vehicle ranking shows some instabilities, particularly in the 16-26 segment, where the risk of 2018 is significantly higher than the one recorded in 2019. Whether to accept this time instability on the model or to decide to remove this variable from the modeling context is something that should be chosen by the modeller.

In this figure, we can see that the claim_history variable is stable across time, as no significant interaction was detected.

In the same modeling project, many different models can be tagged and inspected: for example, to see the differences between a simple model with few variables and a complex model. It is then key to be able to quickly inspect the differences between them. Akur8 gives fast and intuitive comparison thanks to its model comparer.

You will be able to compare all the graphs which are available in the tool: spread, lorenz curve, lift curve and variables.

VARIABLES SPREAD

The spread of a variable is computed from its coefficients:

If a variable is not selected in the model, all its coefficients are equal and the spread equals 0.

For instance, in the graph below, the total (100/0) spread of the driver_age variable is 150%, meaning that the difference in risk between the least risky segment of age and the riskiest is two and half. However, the robust (95/5) spread is 45%. This means that a small segment drives the risk of the variable (in this case, young people).

This can be computed from the coefficients of the variable itself:

→ we can compute the Spread 100/0:

READING VARIABLES GRAPHS

The fitted average and observed averaged should, of course, follow each other. However, if the exposure is low for some levels, the average observed values might vary in non-significant ways. In this case, the coefficients will be grouped together and the model predictions (which are robust) will not follow the observed values.

The coefficients represent the impact of the selected variable, while the average predicted value represents the total impact of all the variables present in the model.

LIFT & LORENZ CURVES

The global models metrics can be visualized in the Model Overview screen.

The lift-curve is built by ordering the observation of the validation sets from the lowest prediction to the highest, and aggregating them by bins containing 5% of the Database. For each bin, the average predicted and observed values are represented.

This curve indicates the discriminatory power of the model (difference between the extreme predicted and observed risk).

If they follow each other closely, it means the model is fitting the data well, while a more chaotic curve means the predictions are not tracking the data well.