Hotstar Re-engagement Strategy

To begin, I'd want to explore the historical data for these 500,000 users. This includes their last active date, frequency and duration of past viewing sessions, primary content genres watched (especially Telugu content), subscription tier and history, device usage, and any available demographic information. Understanding what they used to do is key to figuring out why they stopped and how to bring them back.

Proposed Segmentation Methodology:

1. Value-Based Segmentation (Adapted RFM-C):
   • Recency: While all are 6+ months inactive, I'd look at their last active date within the period before they went dormant. Were they active just before the 6-month cutoff, or did their activity drop off much earlier? This gives a gradient of "dormancy depth."
   • Frequency: How often did they historically use Hotstar per week/month when they were active? Were they daily, weekly, or sporadic users?
   • Monetary Value: What was their subscription tier (if applicable)? If they were on an ad-supported tier, what was their estimated ad revenue contribution based on viewing hours? This helps quantify their past financial value.
   • Content Engagement (the 'C'):
     • Telugu Content Affinity Score: Calculated based on the proportion of Telugu content in their viewing history, watch completion rates for Telugu titles, and frequency of watching new Telugu releases when they were active. I can use techniques like weighted averages or even a simple scoring model here.
     • Genre Preference: Within Telugu content, did they prefer movies, series, reality shows, or sports (if Hotstar carries regional sports feeds)?

   By combining these RFM-C dimensions (e.g., using k-means clustering on these normalized scores or by setting rule-based thresholds), we can create actionable segments such as:

   • "High-Value Telugu Loyalists, Recently Dormant": High past monetary value, high Telugu content affinity, became inactive closer to the 6-month mark. These are prime targets.
   • "Frequent Telugu Snackers, Long Dormant": High frequency, moderate Telugu affinity, but inactive for a longer portion of the 6+ months. Might need stronger incentives.
   • "Low-Engagement, Subscription Lapsed": Low past frequency/monetary value, unclear Telugu affinity, likely cancelled subscription (if applicable). Hardest to re-engage.
   • "Non-Telugu Focused, Accidentally Dormant": Users who happened to be tagged as Telugu but their core consumption was non-Telugu. Re-engagement for them needs different messaging.
2. Subscription Status Segmentation:
   • Still Subscribed but Dormant: These are puzzling. Why pay and not use? Re-engagement is critical to prevent eventual churn.
   • Subscription Lapsed/Cancelled: Re-engagement needs to convince them to re-subscribe.

This segmentation allows us to tailor re-engagement messages, offers, and content recommendations more precisely, rather than a one-size-fits-all campaign for 500k users.

Calculating Telugu Content Affinity Score:

I'd construct a composite score based on several weighted factors from their historical viewing data (when they were active):

1. Proportion of Telugu Watch Time: (Total minutes of Telugu content watched) / (Total minutes of all content watched). Higher weight.
2. Number of Unique Telugu Titles Watched: Indicates breadth of interest in Telugu content.
3. Watch Completion Rate for Telugu Titles: Average completion percentage for Telugu movies/series. Higher completion suggests stronger engagement. Moderate weight.
4. Recency of Telugu Consumption: Was Telugu content among the last things they watched before going dormant? Higher weight if recent.
5. Frequency of Watching New Telugu Releases (when active): Did they tend to watch new Telugu movies/shows soon after release?
6. Explicit Signals (if available): Ratings given to Telugu content, Telugu content added to watchlist.

These factors would be normalized and combined, possibly using a simple weighted sum or a more sophisticated approach if we want to identify latent features of Telugu content preference (e.g., using matrix factorization on user-item interaction data for Telugu content). Users could then be bucketed (e.g., High, Medium, Low Telugu Affinity).

Validating the "Missing Telugu Content" Assumption:

You're absolutely right; we can't assume their dormancy is solely due to missing Telugu content. To diagnose this, I'd pursue several lines of inquiry:

1. Cohort Analysis of Churn/Dormancy:
   • Compare the dormancy/churn rates of users with High Telugu Affinity vs. users with Low Telugu Affinity over the past 6-12 months. If high-affinity Telugu users are becoming dormant at a disproportionately higher rate than other language affinity groups, it suggests a problem specific to Telugu content or its audience.
   • Look at dormancy rates before and after major Telugu releases. If a blockbuster Telugu movie launch didn't reactivate previously engaged Telugu users, why not?
2. Correlate Dormancy with Content Catalog Changes:
   • Did the rate of new, high-quality Telugu content addition slow down prior to or during their dormancy period compared to other languages?
   • Did Hotstar lose rights to any popular library of Telugu content?
3. Competitive Analysis & Market Research:
   • Are competitors like Aha, Prime Video, or Zee5 aggressively acquiring exclusive Telugu content or launching major marketing campaigns targeting Telugu audiences during this period?
   • Conduct surveys (potentially with a small, representative sample of these dormant users if contactable and incentivized) asking about their current streaming habits, preferred platforms for Telugu content, and reasons for not using Hotstar. This can directly ask about shifts in entertainment habits.
   • Look for publicly available data or industry reports on OTT market share shifts in Telugu-speaking regions.
4. Product & UX Factors:
   • Analyze if there were any app updates or UI changes that might have negatively impacted the discovery or experience of Telugu content around the time these users started becoming dormant.
   • Check for technical issues or performance degradation disproportionately affecting users who primarily consume Telugu content (e.g., subtitling issues, audio quality).

The re-engagement strategy must be informed by this diagnosis. If the issue is a fundamental shift away from Hotstar for Telugu content due to a stronger competitor offering, simple content notifications might not be enough; a more strategic response concerning content acquisition or exclusive partnerships might be needed.

Prioritization using a Propensity Model:

1. Model Objective: Predict the likelihood of a dormant user returning to active status (e.g., logging in and watching content for X minutes within Y days) if targeted with a re-engagement campaign.

2. Methodology:

• Training Data: Use historical data of previously dormant users (perhaps those inactive for 3-5 months) who were targeted by past re-engagement efforts (if any) or who returned organically. The target variable would be binary: "Returned" (1) or "Did Not Return" (0).
• Model Choice: I'd start with Logistic Regression for a baseline due to its interpretability and ease of implementation. If performance needs improvement and we have sufficient data, I'd explore Gradient Boosting Machines (like XGBoost or LightGBM), which often provide higher accuracy for tabular data.
• Features for the Model:
  • Historical Engagement: Days since last login, average session duration when active, frequency of sessions, content completion rates, number of unique titles watched.
  • Content Preferences: Telugu Content Affinity Score (as discussed), preferred genres, diversity of content watched.
  • Subscription Data: Current subscription status (active/lapsed), past subscription tenure, plan type, payment history.
  • Demographics (if available & ethical): Age, location (though we're focused on Telugu users, location within AP/TG might matter).
  • Device Information: Primary device used (mobile, TV, web).
  • Past Campaign Interaction: Response to previous promotional emails/notifications.

3. Scoring & Prioritization:

• Score all 500,000 dormant users using the trained propensity model. Each user gets a probability score (e.g., 0 to 1) of returning.
• Tiered Approach:
  • Tier 1 (Top 10-20%, e.g., 50k-100k users): Users with the highest propensity scores (e.g., > 0.6 or 0.7, depending on model calibration and desired precision). These would be the primary focus for more intensive/valuable re-engagement offers.
  • Tier 2 (Next 30-40%): Users with moderate propensity scores (e.g., 0.3 - 0.6). Target with lower-cost, broader campaigns.
  • Tier 3 (Bottom 40-50%): Lowest propensity. Might receive minimal, very low-cost outreach, or be excluded initially to conserve resources.
• Combine with Value: Overlay this propensity score with their historical "Monetary Value" score from the RFM-C segmentation. Prioritize users who are BOTH high propensity AND high historical value (e.g., "High-Value Telugu Loyalists with High Propensity to Return"). This ensures we're focusing on reactivating users who are likely to return and be valuable if they do.

For instance, we could create a final priority score = `w1 * Propensity_Score + w2 * Normalized_Historical_Value_Score`. The weights (w1, w2) can be tuned based on business objectives (e.g., more weight on propensity if budget is very tight, more on value if LTV is paramount).

Justification for Gradient Boosting (vs. Logistic Regression & Neural Networks):

• vs. Logistic Regression (LR):
  • Performance: Gradient Boosting Machines (GBMs) like XGBoost or LightGBM generally offer superior predictive accuracy compared to LR, especially with complex, non-linear relationships and interactions between features, which are common in user behavior data.
  • Feature Handling: GBMs handle categorical features and missing values more gracefully than standard LR implementations often do (though LR can be adapted). They are also less sensitive to feature scaling.
  • While LR offers great interpretability with its coefficients, GBMs provide robust feature importance measures, which can still give good insights into what drives re-engagement.
• vs. Neural Networks (NNs):
  • Data Requirements & Complexity: For a dataset of 500,000 users with potentially dozens of engineered features (tabular data), NNs might be overkill. They typically shine with extremely large datasets or unstructured data (images, text). GBMs often perform as well or better on tabular data of this scale with less tuning and computational cost.
  • Interpretability: NNs are generally more "black box" than GBMs. Explaining feature contributions from a complex NN to business stakeholders is harder than using SHAP values or feature importance from a GBM.
  • Development Time: GBMs are often quicker to develop, train, and tune for such problems.
• Why GBM is a good fit here: It strikes a good balance between high performance, ability to capture complex patterns, reasonable interpretability (via feature importance, SHAP), and efficient training on datasets of this size.

Handling Class Imbalance:

You're right, the "Returned" class will likely be much smaller than "Did Not Return." This imbalance can bias the model towards predicting the majority class.

1. Resampling Techniques:
   • Oversampling the Minority Class: Techniques like SMOTE (Synthetic Minority Over-sampling Technique) create synthetic samples of the minority class. This helps the model learn its characteristics better. Random oversampling is simpler but can lead to overfitting.
   • Undersampling the Majority Class: Randomly remove samples from the majority class. This can lead to loss of information if not done carefully. Techniques like Tomek Links or Edited Nearest Neighbours can be more targeted.
   • A combination of SMOTE and an undersampling technique is often effective.
2. Cost-Sensitive Learning / Class Weights:
   • Many algorithms (including XGBoost/LightGBM and Logistic Regression) allow you to assign different weights to classes during training. We would assign a higher weight to misclassifying the minority "Returned" class, forcing the model to pay more attention to it. The weights can be set inversely proportional to class frequencies.
3. Using Appropriate Evaluation Metrics:
   • Accuracy is Misleading: With high imbalance, a model predicting "Did Not Return" for everyone can achieve high accuracy.
   • Focus on Precision, Recall, F1-Score for the Minority Class:
     • Precision (Positive Predictive Value): Of those predicted to return, how many actually did? `TP / (TP + FP)`. Important for campaign ROI – we don't want to waste effort on false positives.
     • Recall (Sensitivity, True Positive Rate): Of those who actually returned, how many did we predict? `TP / (TP + FN)`. Important for not missing out on potential returners.
     • F1-Score: Harmonic mean of Precision and Recall.
   • Area Under the Precision-Recall Curve (AUC-PR): This is a better metric than ROC-AUC for imbalanced datasets because ROC-AUC can be overly optimistic.
   • Confusion Matrix: To understand the types of errors (FP, FN).
4. Threshold Tuning: The default 0.5 probability threshold for classification might not be optimal. We can adjust this threshold based on the precision-recall trade-off to maximize F1-score or achieve a desired level of precision or recall depending on business goals (e.g., if campaign cost is high, prioritize precision).

I would likely start with class weighting or SMOTE, and heavily rely on AUC-PR and F1-score for model evaluation and selection.

Re-engagement Strategy for Top 100K Users:

1. Campaign Channels & Content (A/B Test Different Combinations):

• Personalized Emails:
  • Subject lines: "We Miss You, [User Name]! Here's What's New in Telugu on Hotstar" or "Remember [Last Watched Major Telugu Title]? You'll Love This..."
  • Content:
    • Highlight 2-3 major Telugu movies/originals they missed, tailored to their past genre preferences (from Content Affinity Score).
    • Showcase any new app features or improvements.
    • Offer a "Welcome Back" incentive (A/B test this): e.g., a 25% discount on their next month if they re-subscribe (if lapsed), or a voucher for a free premium movie rental (if that feature exists).
• Push Notifications (if opted-in and app still installed):
  • Shorter, punchier messages: "🔥 [New Blockbuster Telugu Movie] is now streaming on Hotstar!" or "[Your Favorite Telugu Actor]'s new series just dropped!"
  • Deep-link directly to the content or a personalized "For You" section.
• SMS (use sparingly, for high-value offers if other channels fail):
  • Brief message with a compelling offer and a link.
• (Optional, Higher Cost) Targeted Social Media Ads:
  • Use custom audiences on Facebook/Instagram to target these specific dormant users with ads showcasing new Telugu content.

2. A/B Testing Framework for Campaigns:

• Control Group: A small percentage of the 100K users receive no re-engagement communication (to measure organic return rate).
• Treatment Groups: Test different combinations:
  • Email only vs. Push only vs. Email + Push.
  • Offer vs. No Offer in emails.
  • Different content highlights (e.g., focusing on one blockbuster vs. a variety).

Measuring Success Beyond Initial Login: A Funnel Approach

A successful re-engagement isn't just a login; it's sustained activity and value. I'd track a funnel:

1. Reach & Engagement with Campaign:
   • Email: Open Rate, Click-Through Rate (CTR).
   • Push: Delivery Rate, Open Rate (CTR).
2. Initial Reactivation:
   • App Open Rate / Login Rate: % of targeted users who opened the app or logged in within 7 days of campaign.
   • Subscription Reactivation Rate (if applicable): % of lapsed subscribers who re-subscribe.
3. Post-Reactivation Content Engagement (Crucial):
   • First Meaningful Interaction: % of reactivated users who watch at least X minutes of content (e.g., >10 mins) within their first session or first 3 days. This filters out "curiosity logins."
   • Viewing Hours (VH) per Reactivated User: Average VH in the first week and first month after return.
   • Content Diversity: Number of unique titles/genres watched by reactivated users.
   • Completion Rates for Key Content: Are they finishing the movies/episodes they start?
4. Sustained Retention & Value:
   • 30-Day Active Retention Rate: Of users who reactivated, what % are still active (e.g., logged in and watched content) 30 days later? Compare this to organic returners or new users.
   • 90-Day Active Retention Rate.
   • Resubscription Tenure (if applicable): How long do reactivated subscribers stay subscribed?
   • Incremental LTV (or LTV Recovery): Estimate the LTV of these reactivated users over the next 6-12 months, minus the cost of the re-engagement campaign. This is the ultimate ROI. Compare this to the LTV of a control group of dormant users who weren't targeted or returned organically.
5. Cost per Reactivated User (meaningfully engaged).

Success would be defined by achieving target rates in this funnel, particularly strong post-reactivation engagement and sustained retention, leading to a positive incremental LTV.

Personalization Strategy for Dormant Users (Cool-Start):

My goal is to provide recommendations that are relevant enough to spark interest without relying too heavily on potentially outdated preferences.

1. Leverage Last Known Good State (with caution):
   • Historical Affinity: Use their previously calculated Telugu Content Affinity Score and preferred genres as a starting point, but with a "decay factor." The older the data, the less weight it gets.
   • Last Watched Content: If they were in the middle of a series or watched a specific actor's movie before going dormant, this could be a hook. E.g., "Continue watching [Series Name]?" or "New movie from [Actor of last movie watched]!" – but test this, as they might have abandoned it for a reason.
2. Combine with Popularity & Trends (Content-Based & Global Context):
   • Universally Popular Telugu Content: Feature 1-2 of the absolute biggest Telugu blockbusters or most critically acclaimed/talked-about originals released on Hotstar during their 6-month absence. These have broad appeal and are strong "FOMO" (Fear Of Missing Out) drivers.
   • Trending Content within their Historical Genres: If they liked Telugu action movies, show them the top trending new Telugu action movies, even if not perfectly aligned with their niche past preferences.
   • "What's New & Hot in Telugu": A dedicated section or email highlight of fresh, popular arrivals.
3. Demographic / Segment-Based Recommendations (if affinity data is too sparse/old):
   • If we have reliable demographic segments (e.g., "Young Adults in Hyderabad interested in Action"), recommend content popular within that broader segment. This is less personalized but better than random.
4. Explicit Feedback Loop on Return (Rapid Re-learning):
   • Initial Survey/Preference Selection: Upon first login after a long dormancy, briefly ask them to pick a few genres or titles they are interested in now. "Welcome back! What are you in the mood for?"
   • Interactive Onboarding for Recommendations: Show a few diverse popular titles and let them quickly indicate interest (thumbs up/down, add to watchlist).
   • Prioritize Implicit Signals: Heavily weigh their first few interactions after returning (what they click on, watch, complete) to rapidly update their preference profile. The system should learn quickly from their new behavior.
5. Keep it Simple & Broad Initially for Outreach:
   • For the initial re-engagement email/push, it's safer to highlight 1-2 universally appealing blockbuster Telugu titles they missed rather than trying to guess very niche preferences. The goal of the outreach is to get them back in the app; detailed personalization can then kick in.
6. A/B Test Recommendation Strategies:
   • Test different cool-start recommendation approaches: e.g., Purely Popular vs. Historical Affinity Based vs. Hybrid.

The key is to balance leveraging what little we know from their past with the strong signals of current popular content, and then rapidly adapt based on their behavior once they return.

Pivoting & Decision to Stop Investment:

1. Predefined Failure Thresholds (Set Before Campaign Launch):

• Reactivation Rate: E.g., If < 3-5% of the targeted 100K users reactivate (login + meaningful interaction) within 4 weeks.
• Post-Reactivation 30-Day Retention: E.g., If < 20-30% of the reactivated users are still active after 30 days.
• Cost per Quality Reactivation: If (Campaign Cost / Number of Retained Reactivated Users) significantly exceeds the estimated short-term LTV or a predefined budget per user.

2. If Initial Campaign Fails (Misses Thresholds):

• Immediate Post-Mortem Analysis:
  • Deep Dive into Responders (even if few): Who did respond? Were there any sub-segments within the 100K that showed better (even if still low) response? What were their characteristics? This might inform a much smaller, highly targeted follow-up.
  • Analyze Campaign Execution: Were there technical issues with email delivery, push notifications, broken links? Was the messaging off? Were the offers not compelling enough?
  • Re-evaluate Propensity Model: Were the features or assumptions in the propensity model flawed? Did it incorrectly identify these users as "high propensity"?
  • Qualitative Feedback (if possible): If any users clicked but didn't convert, or if we can survey a small sample, try to understand why the campaign didn't resonate.
• Pivot Strategy Options:
  • Drastically Different Offer/Messaging: If the diagnosis suggests the offer was weak or messaging poor, one final, highly distinct A/B test on a very small sub-segment of the remaining high-propensity users could be tried. E.g., A very deep discount, or highlighting a completely different type of content. This would be a last-ditch, low-cost experiment.
  • Focus on a "Win-Back Window": Shift focus to users who have been dormant for a shorter period (e.g., 1-3 months) where re-engagement likelihood is typically higher, using learnings from the failed campaign.
  • Re-evaluate the Definition of "Dormant": Perhaps 6 months is too long for this specific market/content type, and true churn happens much earlier.

3. Deciding to Stop Investing in this 6-Month+ Dormant Cohort:

• Economic Viability: If even the most optimistic pivot scenarios show a Cost per Quality Reactivation that is higher than their likely recovered LTV, it's economically unviable to continue.
• Opportunity Cost: The resources (time, budget, engineering effort) spent trying to reactivate this deeply dormant base could likely yield a higher return if invested in retaining currently active users or acquiring new, high-potential users.
• Strategic Shift: If diagnosis points to fundamental reasons for dormancy (e.g., massive shift to a competitor with superior exclusive content for Telugu audiences), reactivating them on Hotstar might be fighting an uphill battle. The company might need to address the root cause (e.g., content strategy) rather than just re-engagement tactics.
• Data-Driven Cut-off: After 1-2 failed iterations on this cohort with clear negative ROI, I would recommend officially classifying these users as "Churned - Low Reactivation Potential" and largely cease active re-engagement efforts for them. They might receive very infrequent, extremely low-cost automated "we miss you" emails (e.g., once every 6 months) but no significant resources would be allocated.

The decision to stop is as important as the decision to start. It needs to be data-driven, considering both the direct campaign ROI and the opportunity cost of those resources.

Simplified, Low-Cost Pilot Re-engagement Strategy:

1. Sharper, Simpler Segmentation (Rule-Based):

• Instead of a complex propensity model, use a few clear, rule-based segments based on readily available data:
  • Segment A: "Recent High Telugu Engagers" (e.g., top 20% by Telugu watch time in the 3 months before going dormant, became dormant 6-7 months ago). These are likely the best bet for a simple campaign.
  • Segment B: "Subscription Lapsed - Moderate Telugu Engagers" (e.g., their subscription ended 4-6 months ago, had decent Telugu viewing).
  • Control Group: A random sample from these segments who receive no communication.
• Focus on a smaller pilot group: Instead of 100k, maybe target 10k-20k users from Segment A initially.

2. Single, Lowest-Cost Channel First: Push Notifications (if available) or Email.

• Push Notifications: If a significant portion of these dormant users still have the app installed and notifications enabled, this is the cheapest and most direct channel.
  • Craft 2-3 distinct messages highlighting 1-2 major universally appealing Telugu blockbuster releases they missed or a very compelling new original.
  • A/B test these simple messages.
• Email: If push reach is low, email is the next best low-cost option.
  • Simple, text-focused email (or very light HTML to avoid spam filters) with clear Call-to-Actions.
  • Highlight the top 1-2 pieces of content. A/B test subject lines and content focus.
• No complex personalization initially: Use the same general "blockbuster" content message for everyone in the pilot segment. The personalization is simply that it's Telugu content for Telugu-tagged users.

3. Simplified Offer (Optional, A/B Test):

• If an offer is considered essential, test a single, straightforward offer: e.g., "Watch [New Movie Title] free for 48 hours" or "Get 1 week free on re-subscribing."
• Compare against a no-offer variant.

4. Simplified Success Metrics:

• Primary:
  • 7-Day Reactivation Rate (App Open / Login).
  • For subscription-lapsed, 7-Day Resubscription Rate.
• Secondary (if easily trackable without much setup):
  • Click-Through Rate on the notification/email.
  • Did they watch the highlighted content if they returned?

5. Learning Objective:

• The main goal of this simplified pilot is to quickly gauge if there's any significant responsiveness from these dormant segments to basic outreach about compelling Telugu content. Is there a pulse?
• If even this very low-cost approach yields a response rate significantly above the control group's organic return rate, it provides justification to invest in more sophisticated modeling and multi-channel campaigns later.
• If it fails (e.g., <1% lift over control), it suggests this cohort is very difficult to reactivate, and further investment should be questioned.

This approach strips down the complexity, focuses on the cheapest channels, uses simpler segmentation, and aims for quick learnings on a smaller scale to inform whether a larger investment is warranted.

Presenting Findings to Non-Technical Stakeholders:

1. Focus on "What Does This Mean for the Business?":

• Start with the Big Picture: "We have a segment of 500,000 dormant Telugu users, representing a potential [X]% of our Telugu user base and an estimated [Y] in potential lost subscription/ad revenue annually if they remain inactive."
• Key Learnings in Simple Terms:
  • "Our analysis shows that users who historically watched a lot of Telugu content [Segment X] are [more/less] likely to return if we remind them about new blockbusters like '[Movie Title]'."
  • "Campaigns featuring [Specific Offer/Content Type] were [Z]% more effective at bringing users back than [Alternative Approach]."
  • "Users who return tend to watch [A] hours in their first month back, which could translate to [B] in recovered LTV."
• Visualizations: Use simple charts and graphs – bar charts for segment sizes, line graphs for return rates over time, funnels for campaign performance. Avoid complex statistical plots.

2. Actionable Recommendations, Not Just Data:

• "Based on these findings, we recommend prioritizing re-engagement efforts on [Specific User Segment] using [Most Effective Campaign Type]."
• "For the next phase, we suggest A/B testing [New Idea] to further improve return rates for [Another Segment]."
• "This could potentially reactivate an additional [N] users, leading to an estimated [₹Y] in incremental revenue over the next year."

3. Highlight ROI and Cost-Effectiveness:

• "Our pilot campaign cost [₹C] and brought back [N] users who are now actively engaging. The cost per reactivated engaged user was [₹C/N], which compares favorably/unfavorably to our target."

Communicating Potential Biases and Limitations:

It's crucial to maintain credibility by being transparent about what the data can and cannot tell us:

1. Correlation vs. Causation:
   • "While our campaign was followed by [N] users returning, we must be cautious. Our control group helps isolate our impact, but external factors (e.g., a competitor's price increase, a viral social media trend about a Hotstar show) could also influence returns. We've tried to account for this, but it's hard to be 100% certain our campaign was the sole cause."
2. Survivorship Bias (in analyzing historical data of dormant users):
   • "When we look at the past behavior of these currently dormant users, we're looking at users who, by definition, eventually became inactive. Their historical patterns might differ from users who remained active. This can affect how well our models generalize."
3. Selection Bias (in who we target or who responds):
   • "If we heavily target users based on our propensity model, those who return might be inherently different from those we didn't target or who didn't respond. The 'average' reactivated user might not represent all 500k dormant users."
   • "Users who respond to emails/pushes might be more digitally savvy or have a lingering affinity for Hotstar, making them easier to win back than the broader dormant group."
4. Staleness of Data & Changing Preferences:
   • "User preferences can change significantly in 6 months. Our recommendations based on past behavior are an educated guess. The 'cool-start' recommendations for returning users are designed to quickly relearn their current interests."
5. Short-Term vs. Long-Term Effects:
   • "Our initial results show [X] reactivation. However, we need to monitor these users for several months to understand if this re-engagement is sustained or if they quickly become dormant again. True success is long-term retention."
6. Model Limitations:
   • "Our propensity model predicts likelihood based on past data. It's not a crystal ball and will have false positives (predicting return for users who don't) and false negatives (missing users who might have returned with a different nudge)." We aim to optimize this trade-off.

By framing insights around business impact and transparently discussing limitations, we can build trust and facilitate better data-informed decision-making by the stakeholders.