NUS-SDS-Datathon-Singlife

Dataset Handling Guidelines

• The dataset is provided by Singlife. The dataset is strictly for the purpose of this competition and should not be used for any other purposes.
• The dataset must not be shared with anyone outside of the competition, or uploaded to any repositories / cloud storage.

GitHub Repository Instructions

• The forked repository is public by default and it can't be made private. If you wish to collaborate on the project through GitHub, you can create a private repository and add your teammates as collaborators.
• Here is a set of instructions to create a private mirror of the repository.
• Make sure to make your repository public for your final submission.

Column Header Descriptions

General Client Information

• clntnum: Unique identifier for the client.
• race_desc: Description of the client's race.
• ctrycode_desc: Country code indicating the client's location.
• clttype: Customer status.
• stat_flag: Flag indicating ACTIVE, LAPSED or MATURED. E.g. if there’s at least one inforce policy, then the flag would be ACTIVE. If all of the client’s policies are all lapsed, then it is LAPSED.
• min_occ_date: Date of the client's first interaction or policy purchase with the company.
• cltdob_fix: Fixed or corrected date of birth of the client.
• cltsex_fix: Fixed or corrected gender of the client.

Client Risk and Status Indicators

• flg_substandard: Flag for substandard risk clients.
• flg_is_borderline_standard: Flag for borderline standard risk clients.
• flg_is_revised_term: Flag if customer ever has revised terms.
• flg_is_rental_flat: Indicates if the client lives in a rental flat.
• flg_has_health_claim: Flag for clients with health insurance claims.
• flg_has_life_claim: Flag for clients with life insurance claims.
• flg_gi_claim: Flag for general insurance claims.
• flg_is_proposal: Indicates if there is a policy in proposal for client.
• flg_with_preauthorisation: Flag for clients with preauthorized transactions or policies.
• flg_is_returned_mail: Flag for returned mail instances.

Client Consent and Communication Preferences

• is_consent_to_mail, is_consent_to_email, is_consent_to_call, is_consent_to_sms: Flags indicating client's consent to various forms of communication.
• is_valid_dm, is_valid_email: Flags indicating the validity of direct mail and email addresses.

Demographic and Household Information

• is_housewife_retiree, is_sg_pr, is_class_1_2: Flags indicating specific demographics like occupation, residency status, etc.
• is_dependent_in_at_least_1_policy: Indicates if the client is a dependent in at least one policy.
• hh_20, pop_20, hh_size, hh_size_est: Metrics related to household size and population.
• annual_income_est: Estimated annual income of the client, in buckets.

Policy and Claim History

• n_months_last_bought_products, flg_latest_being_lapse, flg_latest_being_cancel, recency_lapse, recency_cancel: Metrics related to the recency of policy purchases, lapses, and cancellations.
• tot_inforce_pols, tot_cancel_pols: Total number of in-force and canceled policies.
• f_ever_declined_la: Flag for clients has ever been declined policies.

Anonymized Insurance Product Metrics (APE, Sum Insured, Prepaid Premiums)

• ape_, sumins_, prempaid_* (e.g., ape_gi_42e115, sumins_ltc_1280bf, prempaid_grp_6fc3e6): Metrics for various anonymized insurance products, likely representing different types of policies like general insurance, long-term care, group policies, etc. The suffixes (like 42e115, 1280bf) are unique identifiers for the specific insurance products. ‘ape’ stands for Annual Premium Equivalent, 'sumins' for sum insured, ‘prempaid’ stands for premium customers will pay from product inception to product maturity.

Other Flags and Metrics

• f_elx, f_mindef_mha, f_retail: Flags possibly related to client's association with specific programs or sectors.
• flg_affconnect_*, affcon_visit_days, n_months_since_visit_affcon: Metrics related to client’s activity in affinity connect.
• clmcon_visit_days, recency_clmcon, recency_clmcon_regis: Metrics related to client’s activity in claim connect.
• hlthclaim_amt, giclaim_amt, recency_hlthclaim, recency_giclaim, hlthclaim_cnt_success, giclaim_cnt_success: Health and general insurance claim-related metrics.
• flg_hlthclaim_, flg_gi_claim_ (e.g., flg_hlthclaim_839f8a_ever, flg_gi_claim_29d435_ever): Flags for specific types of health and general insurance claims, with anonymized identifiers.

Purchase and Lapse Metrics for Specific Products

• f_ever_bought_, n_months_last_bought_, lapse_ape_, n_months_since_lapse_ (e.g., f_ever_bought_839f8a, n_months_last_bought_grp_6fc3e6, lapse_ape_ltc_1280bf, n_months_since_lapse_inv_dcd836): Flags and metrics indicating purchase history, lapses, and time since last interaction for various anonymized insurance products.

Target Column

• f_purchase_lh: Flag indicating whether a customer will purchase life or health insurance products within the next three months.

Submission Instructions

1. Rename the submission notebook as "NUS_DATATHON_SINGLIFE_<TEAM NAME>.ipynb". For example, if your team name is "NUS 1", your submission notebook should be named as "NUS_DATATHON_SINGLIFE_NUS 1.ipynb".
   The notebook already contains a template function, named testing_hidden_data, to take in the test data and output the predictions. You are to fill in the function with your model and any preprocessing function that you may have.
   Note: Each team should only submit one notebook and the notebook should be in the master branch.

2. You can save your model and upload it in the GitHub repository provided that it does not exceed the file size limit on GitHub (see Section 3 on instructions for large models). If you saved your model, you are to provide a script the load the model. Please name the loading function as load_model.
   If the model is not saved, we will assume that the model is trained using the training function and the output of the training model is a model that can be is trained using the template function, testing_hidden_data, provided in the notebook. Please name the training function as train_model.

3. If your model is too large to be uploaded to GitHub, you can upload it to Google Drive and share the link with us. Please ensure that the link is accessible by anyone with the link. Please provide a LOAD_MODEL_INSTRUCTIONS.md file in the GitHub repository to explain how to load the model. Note that you are strongly discouraged to use such large models.