Predict Live-Service Game Retention | Dataford Interview Questions

Business Context

NovaForge Studios operates a live-service mobile RPG with 8M monthly active players. The growth team wants a model that predicts whether a new player will still be active on day 7 so they can trigger CRM campaigns, in-game offers, and onboarding interventions.

Dataset

The training data is built at the player-install level using behavior from the first 48 hours after install. The label is whether the player is retained on day 7.

Feature Group	Count	Examples
Session behavior	14	session_count_48h, total_minutes_48h, avg_session_length, last_session_gap_hours
Progression	10	levels_completed, tutorial_completion_pct, deaths_per_level, quest_completion_rate
Economy & monetization	8	soft_currency_earned, hard_currency_spent, first_purchase_flag, ad_views
Social & engagement	6	guild_joined, friends_added, chat_messages, push_opt_in
Device & acquisition	7	platform, country_tier, install_source, device_ram_gb
Temporal features	5	install_hour_local, weekend_install, day1_day2_play_delta, recency_hours

Size: 1.2M player installs over 6 months, 50 features
Target: Binary — retained on day 7 (1) vs not retained (0)
Class balance: 28% retained, 72% churned before day 7
Missing data: ~12% missing in acquisition metadata, ~6% missing in device specs, sparse missingness in social features for players who never unlocked them

Success Criteria

A good solution should achieve strong ranking quality for intervention targeting, with ROC-AUC >= 0.82, PR-AUC >= 0.58, and top-decile lift >= 2.2 on a strictly future holdout set.

Constraints

Predictions must be generated in a daily batch for up to 500K new installs
Marketing and game design need feature-level explanations for why players are flagged
The model should be robust to changing acquisition mix and live-ops events
Retraining budget supports weekly retraining, not continuous online learning

Deliverables

Define the target and leakage-safe feature window
Build and justify a retention prediction model
Describe preprocessing and feature engineering choices
Evaluate the model with business-relevant metrics and thresholding
Explain how you would deploy, monitor, and retrain the model

Problem

Business Context

Dataset

The training data is built at the player-install level using behavior from the first 48 hours after install. The label is whether the player is retained on day 7.

Feature Group	Count	Examples
Session behavior	14	session_count_48h, total_minutes_48h, avg_session_length, last_session_gap_hours
Progression	10	levels_completed, tutorial_completion_pct, deaths_per_level, quest_completion_rate
Economy & monetization	8	soft_currency_earned, hard_currency_spent, first_purchase_flag, ad_views
Social & engagement	6	guild_joined, friends_added, chat_messages, push_opt_in
Device & acquisition	7	platform, country_tier, install_source, device_ram_gb
Temporal features	5	install_hour_local, weekend_install, day1_day2_play_delta, recency_hours

Size: 1.2M player installs over 6 months, 50 features
Target: Binary — retained on day 7 (1) vs not retained (0)
Class balance: 28% retained, 72% churned before day 7
Missing data: ~12% missing in acquisition metadata, ~6% missing in device specs, sparse missingness in social features for players who never unlocked them

Success Criteria

A good solution should achieve strong ranking quality for intervention targeting, with ROC-AUC >= 0.82, PR-AUC >= 0.58, and top-decile lift >= 2.2 on a strictly future holdout set.

Constraints

Predictions must be generated in a daily batch for up to 500K new installs
Marketing and game design need feature-level explanations for why players are flagged
The model should be robust to changing acquisition mix and live-ops events
Retraining budget supports weekly retraining, not continuous online learning

Deliverables

Define the target and leakage-safe feature window
Build and justify a retention prediction model
Describe preprocessing and feature engineering choices
Evaluate the model with business-relevant metrics and thresholding
Explain how you would deploy, monitor, and retrain the model

Problem

Business Context

Dataset

The training data is built at the player-install level using behavior from the first 48 hours after install. The label is whether the player is retained on day 7.

Feature Group	Count	Examples
Session behavior	14	session_count_48h, total_minutes_48h, avg_session_length, last_session_gap_hours
Progression	10	levels_completed, tutorial_completion_pct, deaths_per_level, quest_completion_rate
Economy & monetization	8	soft_currency_earned, hard_currency_spent, first_purchase_flag, ad_views
Social & engagement	6	guild_joined, friends_added, chat_messages, push_opt_in
Device & acquisition	7	platform, country_tier, install_source, device_ram_gb
Temporal features	5	install_hour_local, weekend_install, day1_day2_play_delta, recency_hours

Size: 1.2M player installs over 6 months, 50 features
Target: Binary — retained on day 7 (1) vs not retained (0)
Class balance: 28% retained, 72% churned before day 7
Missing data: ~12% missing in acquisition metadata, ~6% missing in device specs, sparse missingness in social features for players who never unlocked them

Success Criteria

A good solution should achieve strong ranking quality for intervention targeting, with ROC-AUC >= 0.82, PR-AUC >= 0.58, and top-decile lift >= 2.2 on a strictly future holdout set.

Constraints

Predictions must be generated in a daily batch for up to 500K new installs
Marketing and game design need feature-level explanations for why players are flagged
The model should be robust to changing acquisition mix and live-ops events
Retraining budget supports weekly retraining, not continuous online learning

Deliverables

Define the target and leakage-safe feature window
Build and justify a retention prediction model
Describe preprocessing and feature engineering choices
Evaluate the model with business-relevant metrics and thresholding
Explain how you would deploy, monitor, and retrain the model

Problem

Business Context

Dataset

The training data is built at the player-install level using behavior from the first 48 hours after install. The label is whether the player is retained on day 7.

Feature Group	Count	Examples
Session behavior	14	session_count_48h, total_minutes_48h, avg_session_length, last_session_gap_hours
Progression	10	levels_completed, tutorial_completion_pct, deaths_per_level, quest_completion_rate
Economy & monetization	8	soft_currency_earned, hard_currency_spent, first_purchase_flag, ad_views
Social & engagement	6	guild_joined, friends_added, chat_messages, push_opt_in
Device & acquisition	7	platform, country_tier, install_source, device_ram_gb
Temporal features	5	install_hour_local, weekend_install, day1_day2_play_delta, recency_hours

Size: 1.2M player installs over 6 months, 50 features
Target: Binary — retained on day 7 (1) vs not retained (0)
Class balance: 28% retained, 72% churned before day 7
Missing data: ~12% missing in acquisition metadata, ~6% missing in device specs, sparse missingness in social features for players who never unlocked them

Success Criteria

A good solution should achieve strong ranking quality for intervention targeting, with ROC-AUC >= 0.82, PR-AUC >= 0.58, and top-decile lift >= 2.2 on a strictly future holdout set.

Constraints

Predictions must be generated in a daily batch for up to 500K new installs
Marketing and game design need feature-level explanations for why players are flagged
The model should be robust to changing acquisition mix and live-ops events
Retraining budget supports weekly retraining, not continuous online learning

Deliverables

Define the target and leakage-safe feature window
Build and justify a retention prediction model
Describe preprocessing and feature engineering choices
Evaluate the model with business-relevant metrics and thresholding
Explain how you would deploy, monitor, and retrain the model