What is a cohort in mobile gaming UA?

A cohort is a group of users who installed your game within the same defined time window, typically a day, a week, or a campaign flight. Cohort analysis tracks how that specific group behaves and generates revenue over time, isolating them from users acquired in different periods. This is the foundation of LTV forecasting: rather than looking at all active users, you look at each cohort's maturation curve individually.

How many installs do you need before cohort forecasting is reliable?

For reliable D365 predictions at campaign level using ML-based cohort forecasting trained on cross-customer aggregates, the minimum is roughly 500 installs per cohort window, with 2,000+ preferred. Pure single-studio curve-fitting models without cross-customer training data require substantially more data depth to achieve comparable confidence, as Eric Seufert documents in Mobile Dev Memo's classic analysis of LTV data requirements. Studios running very segmented UA (small cohorts per channel and geo) can aggregate to a higher level of analysis to maintain reliability.

How does cohort forecasting handle seasonality?

Production cohort models incorporate seasonal controls, distinguishing between Q4 cohorts (higher early revenue due to holiday spending) and Q1 cohorts (typically lower). This prevents the common forecasting error of over-projecting Q1 performance based on Q4 training data. Models without seasonal controls systematically misforecast for the first 12 months after deployment.

Can cohort forecasting work for free-to-play games with very long LTV tails?

Yes, and these games are where cohort forecasting adds the most value. Strategy games, 4X titles, and RPGs often have users who generate meaningful revenue at D180 to D365 that would not appear in a standard 30-day analysis. Long-horizon cohort models are specifically designed to capture these whale tails accurately, where simpler curve-fitting approaches break down.

How does cohort forecasting handle games with both IAP and IAA revenue?

Hybrid monetisation models require blended LTV predictions that account for both in-app purchase revenue and in-app advertising revenue from the same cohort. A user who makes no purchases may still generate substantial D365 IAA revenue. Pure-IAP models consistently understate LTV for hybrid titles. Well-built cohort forecasting pipelines model both revenue streams separately and combine them into a blended LTV curve.

What is the difference between LTV forecasting and cohort forecasting?

LTV forecasting typically refers to the predicted revenue of an individual user or user segment. Cohort forecasting is the broader discipline, predicting the behaviour of an entire acquisition cohort, including retention, ARPU trajectory, and campaign ROAS. Cohort-level forecasts are generally more stable than individual-user models because they aggregate away the noise of single-user variance, while still preserving the structural signal of the acquisition window.

How often do cohort forecasts update?

Production cohort forecasting platforms refresh predictions daily as new revenue data arrives from the MMP. Early predictions (D1 to D7) are flagged as higher uncertainty. By D14 most models have converged to within their final confidence interval.

Can cohort forecasting be used for investor reporting and M&A?

Yes, and this is one of the highest-value use cases. Investors and acquirers increasingly require cohort-based LTV projections as part of due diligence. Standardised cohort waterfall reports showing D30, D90, D180, D365 ARPU by acquisition period are the format that leading gaming investors and M&A advisors recognise and trust. Studios with rigorous cohort forecasting can compress diligence timelines and command higher multiples because the forward revenue case is quantified rather than asserted.

Does cohort forecasting work with all major MMPs?

Most modern cohort forecasting platforms integrate natively with AppsFlyer and Singular, with Adjust support common but sometimes lagging on feature parity. Custom CSV ingestion or warehouse-direct integrations are available for studios using other attribution solutions or proprietary data pipelines.

How does cohort forecasting change with ATT and privacy-first attribution?

Privacy changes have reduced user-level signal quality, but cohort-level modelling is more resilient than individual attribution. Cohort models operate on aggregate cohort data rather than user-level device IDs, making them structurally compatible with ATT opt-out environments, SKAdNetwork reporting, and Google's Privacy Sandbox. The studios most affected by privacy changes were those relying on user-level retargeting and last-click attribution; those running cohort-based LTV forecasting have been more resilient.

Should a studio build cohort forecasting in-house or use a specialist platform?

For most studios, the answer is to use a specialist platform for the foundational LTV prediction layer and augment with in-house tooling for studio-specific extensions. Specialist platforms typically deliver forecasts in weeks, where an equivalent in-house build runs 6 to 18 months and $1.5M+ per year in fully-loaded data science headcount before producing trustworthy output. Pure in-house builds are justified only when the studio's monetisation model is unique enough that off-the-shelf models cannot calibrate well, or when the studio treats cohort forecasting as a core strategic differentiator.

How accurate are cohort forecasts in practice?

Accuracy varies by method and training data depth. In Kohort's back-testing across $6B+ of UA spend, simple linear extrapolation from D30 data reaches roughly 70% accuracy at D365, regression models on studio-specific data reach 75 to 82%, and gradient-boosted ML models trained on 6+ months of mobile gaming cohort data consistently achieve 88 to 92% accuracy at D365. The gap widens for games with heavy long-tail LTV, where linear extrapolation breaks down structurally.

How do confidence intervals work in cohort forecasting?

A confidence interval expresses the range of plausible outcomes around a point forecast, typically at 80% or 95% probability. A D365 ARPU forecast of $5.00 with a plus or minus 10% interval at 90% confidence means there is a 90% probability the realised D365 ARPU will fall between $4.50 and $5.50. Confidence intervals narrow as more cohort data arrives and as training data depth grows.

Guide

What is Cohort Forecasting? A Complete Guide for Mobile Game Studios

How modern mobile game studios use machine learning to predict cohort revenue, retention, and lifetime value weeks or months before the data has matured. Estimated reading time: 14 minutes.

TL;DR

What you will learn on this page

This guide covers the full discipline of cohort forecasting for mobile game studios: what it is, how it works step by step, the main methods used to forecast cohort LTV, the most common mistakes that produce unreliable forecasts, the landscape of tools mature studios use, and a plain-English glossary of the metrics that matter. If you run UA, finance, or analytics at a mobile studio and your forward LTV decisions cost more than $50K per month to get wrong, this page is written for you.

Definition

What is Cohort Forecasting?

Cohort forecasting is the practice of grouping users acquired in the same time window (a "cohort") and predicting that group's future revenue, retention, and lifetime value using machine learning models trained on historical cohort behaviour, producing a forward view of profitability weeks or months before the data has matured.

The four pillars of cohort forecasting in mobile gaming:

Cohort definition

Grouping installs by acquisition date, channel, geo, or creative to isolate comparable user groups so they can be compared apples to apples.

LTV curve fitting

Modelling the revenue trajectory from Day 0 to D365+ based on early-day signals (retention, session depth, early in-app purchase events).

Prediction confidence

Quantifying uncertainty so decision makers know when a forecast is actionable versus when to wait for more data to accumulate.

Campaign attribution

Mapping cohort revenue back to specific campaigns, creatives, and channels so the forecast feeds budget decisions, not just retrospective reporting.

The fundamental challenge in mobile gaming is that you spend money today to acquire users whose full revenue value will not be known for 6 to 18 months. Cohort forecasting compresses that gap by using the early behaviour of a new cohort, first session length, D1/D3/D7 retention, early IAP signals, to predict how that cohort will perform at D90, D180, and D365.

The discipline draws on decades of work in customer cohort analysis from direct response and SaaS marketing, adapted to the specific shape of mobile gaming revenue curves: high variance, heavy-tailed distributions, IAP plus IAA hybrid streams, and the privacy-first attribution environment introduced by ATT and SKAdNetwork. It has become the primary competitive lever for mobile UA, finance, and M&A teams who need to make forward-looking budget and valuation decisions on cohorts that have not yet matured.

How it works

How Cohort Forecasting Works

Most studios still report cohort performance only after the fact, using trailing 30, 60, or 90 day windows. By the time those windows close, budget has been committed. Modern cohort forecasting inverts the workflow: forward predictions generated in the first week of a cohort, refined daily as more data arrives, used as the primary input to active budget and planning decisions.

Step 1

Cohort data ingestion

Cohort forecasting platforms connect to a studio's MMP (AppsFlyer, Singular, or Adjust) and pull structured cohort revenue data: installs, events, and revenue by day, segmented by channel, geo, and creative. Data is normalised and deduplicated to handle the common edge cases in MMP exports (re-attribution windows, view-through attribution, organic bleed-through). Output: clean per-cohort timeseries ready for model training.

Step 2

LTV model training

ML models are trained on the studio's own historical cohort data, not industry averages. LTV curves vary significantly by game genre, monetisation model (IAP vs IAA vs hybrid), and market. The most accurate approaches combine studio-specific training data with cross-customer aggregates as a regularisation signal, especially for new titles where the studio's own training data is thin. Most production-grade platforms use gradient-boosted models with cohort-level features as inputs and D90/D180/D365 ARPU as targets, re-trained automatically as new cohort data arrives.

Step 3

Forecast generation and confidence scoring

For each new cohort (typically defined weekly or per campaign flight), the platform generates a point forecast plus a confidence interval at D90, D180, and D365. Confidence scoring tells UA managers when a prediction is reliable enough to act on. As a rule of thumb in Kohort's models, which use gradient-boosted ML trained on cross-customer cohort aggregates: cohorts with 500+ installs and 7+ days of data typically fall within plus or minus 15% of D365 actual; 2,000+ installs and 14+ days within plus or minus 8%. Below 500 installs, forecasts should be treated as directional rather than precise. These thresholds are tighter than what a single-studio curve-fitting model can achieve at the same data depth, as Eric Seufert documents in Mobile Dev Memo's classic analysis of LTV data requirements. The cross-customer training data is what compresses the data-depth requirement.

Step 4

ROAS mapping and budget signals

Predicted LTV multiplied by cohort size, divided by cohort-level ad spend, produces a predicted campaign ROAS at D90/D180/D365. Cohort forecasting platforms map this back to live campaigns in real time, so UA managers see which campaigns are tracking to target ROAS before the standard trailing window would surface the signal. This is the step that turns a cohort forecast from a reporting artifact into a budget allocation tool, the core feedback loop behind systematic user acquisition optimization.

Step 4 above is where cohort forecasting feeds directly into user acquisition optimization: the same forward signal that ranks cohorts by predicted LTV ranks live campaigns by predicted ROAS.

Methods

Cohort Forecasting Methods Compared

There is no single "right" way to forecast cohort LTV. The right method for a given studio depends on training data depth, technical capability, and the LTV horizon being forecast.

Method	How it works	Best for	Limitations
Linear extrapolation	Multiply D7 or D30 revenue by an industry multiplier to estimate D365	Smallest studios with no in-house analytics; very early launch windows	Roughly 70% accuracy at D365; ignores cohort-specific behaviour; fails on long-tail and hybrid monetisation titles
Parametric curve fitting (power, Weibull, exponential)	Fit a parametric retention or revenue function (Russell Ovans' r(n)=a·n^b power function, Weibull decay, exponential, or the logarithmic "Golden Curve") to observed cohort data and project forward	Studios at any scale with clean retention data; the most widely used production approach for retention curve modelling outside large vendor stacks; used by practitioners like Russell Ovans at East Side Games	Curve shape assumption can be miscalibrated for new titles; combining IAP and IAA revenue streams in a single curve loses signal; tail behaviour at D365+ is sensitive to early-day noise
Probabilistic models (Pareto/NBD, BG/NBD)	Model purchase frequency and churn separately using classical probability distributions; widely cited in the customer-analytics literature (Schmittlein, Fader/Hardie)	Studios with strong analyst capacity and a need for interpretable, theory-grounded LTV models; common in academic and finance settings	Assumes purchase behaviour follows specific distributions that may not fit gaming cohorts well; less common in production mobile gaming workflows than curve fitting or ML
Cohort regression	Multivariate regression on cohort features (retention, session count, early ARPU) trained on historical cohorts	Studios with 6+ months of cohort history and analyst time to build and maintain models	Limited capacity to capture non-linear interactions; manual model maintenance
Gradient-boosted ML	Boosted decision trees (XGBoost, LightGBM) on cohort features, retrained automatically as data arrives	Most production-grade cohort forecasting today; balances accuracy with interpretability; particularly strong on high-LTV segments	Requires training data depth (90+ days minimum); benefits significantly from cross-customer aggregates; production setups often pair gradient boosting with Tweedie or zero-inflated regression for low-LTV stability
Deep learning sequence models	LSTM or Transformer architectures on cohort timeseries	Largest studios with extensive ML infrastructure and very long LTV tails	Compute-intensive; harder to interpret; marginal gains over gradient-boosted approaches at most scales

Most current best-in-class platforms use gradient-boosted models as the production layer, often supplemented with parametric priors for sparse-data cohorts. In Kohort's back-testing across $6B+ of UA spend, gradient-boosted models trained on 6+ months of studio-specific cohort data achieve D365 prediction accuracy in the 88 to 92% range on cohorts of sufficient size, where simple linear extrapolation from D30 data sits closer to 70%. The gap widens for games with heavy long-tail LTV, where linear extrapolation breaks down structurally.

Pitfalls

Common Mistakes in Cohort Forecasting

The following six mistakes account for the largest share of bad cohort forecasts in mobile gaming.

Mistake 1

Aggregating cohorts too broadly

Studios that run cohort forecasting at the portfolio level rather than at the campaign or channel level lose most of the signal. A 'May 2026 cohort' containing installs from Meta, Google UAC, and AppLovin will have a blended LTV curve that hides material variance between sources. The fix is forecasting at the smallest cohort definition that has statistical significance (500+ installs minimum), then rolling up.

Mistake 2

Comparing predicted LTV to incomplete actual LTV

A common reporting error: comparing a model's D365 prediction against a cohort's current D60 actual revenue, and concluding the model is 'wrong' because the actual is lower. The actual will catch up. The right comparison is predicted D-N versus actual at the same horizon (predicted D365 against actual D365), which requires waiting for cohorts to mature before judging model accuracy. This is one of the most common reasons studios lose faith in cohort forecasting prematurely.

Mistake 3

Ignoring confidence intervals when making budget decisions

A point forecast with a plus or minus 25% confidence band is qualitatively different from a forecast with a plus or minus 5% band, even if the central estimate is the same. Studios that treat all forecasts as equally precise will scale campaigns based on noisy signal in the early weeks of a cohort and have to walk it back. The fix is requiring a minimum confidence threshold (typically plus or minus 15% or tighter) before letting forecasts drive material budget shifts.

Mistake 4

Conflating retention and revenue curves

Cohort retention (the share of users still active at day N) and cohort revenue (ARPU at day N) are related but not the same. Two cohorts can have identical D30 retention and very different D30 revenue, depending on monetisation depth. Forecasting only on retention misses the monetisation signal; forecasting only on revenue misses early dropout. Mature cohort models use both as input features.

Mistake 5

Not retraining as the game economy changes

A cohort model trained on data from before a major monetisation update (new IAP currency, ad placement change, live ops cadence shift) will systematically misforecast cohorts acquired after the update. Production cohort forecasting requires automatic retraining as new cohort data arrives, plus the ability to detect and flag drift when actuals diverge from predictions in a structured way.

Mistake 6

Using cross-genre benchmarks as a forecast input

LTV curves vary dramatically by genre: hypercasual games monetise heavily in the first 7 days primarily through IAA, while 4X and strategy games carry long, heavy IAP-driven tails where a meaningful share of D365 LTV materialises after Day 90, in some sub-genres the majority. Studios using generic mobile gaming benchmarks to evaluate a specific genre will systematically over-bid on slow-monetising titles and under-bid on fast-monetising ones. The fix is genre-calibrated training data, either from the studio's own portfolio or from a cross-customer aggregate filtered to comparable titles.

Landscape

The Cohort Forecasting Tool Landscape

Cohort forecasting draws from several categories of tooling, and most mature studios use a combination.

Mobile Measurement Partners (MMPs)

AppsFlyer, Singular, Adjust, and Kochava are the data foundation for cohort forecasting. They define cohorts at the install level, track in-app events and revenue, and produce the structured timeseries that prediction models consume. MMPs do not themselves produce forward LTV forecasts. They produce the data that forecasting tools predict on.

In-house BI and Data Warehouses

Metabase, Looker, Tableau, and similar BI tools pulling from a centralised data warehouse handle retrospective cohort analysis well. They answer "what was the D90 ARPU of our April cohort?" efficiently. They rarely contain the ML layer needed to answer "what will the D365 ARPU of our May cohort be?"

Statistical and ML Stacks

Studios with in-house data science teams sometimes build cohort forecasting on top of Python (with libraries like lifelines, scikit-learn, XGBoost) or R. The barrier is not the libraries (which are mature) but the productionisation: data pipelines, automatic retraining, drift detection, integration into the UA tooling stack, and ongoing maintenance against MMP API changes.

Specialist Cohort Forecasting Platforms

This category, which includes Kohort's Ktrl as well as tools like GameAnalytics (for smaller studios) and bespoke in-house ML stacks, sits on top of MMP data and adds the forward prediction layer. Differentiators within this category are model accuracy at early prediction windows (D7 and D14), training data approach (industry averages versus studio-specific versus cross-customer aggregate), and how tightly the forecast is integrated into campaign-level ROAS decisions.

Specialist platforms typically deliver production-grade cohort forecasts in weeks. Equivalent in-house builds usually take 6 to 18 months before producing trustworthy output, and run roughly $1M to $2M per year in fully-loaded data science headcount depending on team size and seniority. Current US ML engineer total compensation sits at $160K to $260K per role; a production-grade cohort forecasting pipeline typically requires 2 to 3 dedicated engineers plus shared infrastructure. Specialist platforms also aggregate cross-customer training data, which improves early-window prediction accuracy in ways a model trained only on a single studio's history cannot match.

Most studios spending $5M+ per month on UA, or running M&A processes on a mobile portfolio, run a specialist cohort forecasting platform for the foundational LTV prediction layer, then augment with in-house tooling for studio-specific extensions (custom cohort segmentations, investor reporting templates, integrations into the financial planning stack). The pattern mirrors how the same studios use MMPs like AppsFlyer or Singular rather than building attribution from scratch.

Outcomes

Studios Using Cohort Forecasting: Results

Mobile game studios across hypercasual, casual, mid-core, and hybrid models have implemented cohort forecasting with measurable outcomes.

Lotum (4 Pics 1 Word)

Makers of one of the longest-running casual word games scaled UA spend by 55% and increased revenue by 15% by combining long-tail cohort LTV forecasting with disciplined budget reallocation, capturing meaningful revenue from the D90+ tail that trailing windows had been missing entirely.

Read case study →

52 Entertainment

Global evergreen games leader achieved a 13% ROAS lift in 4 weeks and a 3.5x improvement in campaign efficiency by moving from trailing ROAS windows to forward cohort-based budget allocation across their gaming portfolio.

Read case study →

Apps Teknoloji

Leading Turkish mobile gaming studio reduced campaign analysis time by 30% and saved 5+ hours per week per UA manager by replacing manual LTV prediction workflows with real-time predictive cohort ROAS signals.

Read case study →

See all customer outcomes at /customers.

Glossary

Cohort Forecasting Glossary

The following terms appear throughout this page and across cohort forecasting discussions.

Cohort: A group of users defined by a shared acquisition characteristic, most commonly install date (daily, weekly, or campaign-flight cohorts).
LTV (Lifetime Value): The total revenue a user or cohort is expected to generate over a defined period (D30, D90, D180, D365). Usually expressed as ARPU at a given day horizon.
ARPU (Average Revenue Per User): Total cohort revenue divided by total installs in the cohort, at a specified day horizon (D7 ARPU, D90 ARPU, etc.).
ARPDAU (Average Revenue Per Daily Active User): Total daily revenue divided by daily active users. A different denominator than ARPU; reports on engagement quality rather than cohort-wide value.
Retention curve: The percentage of users still active at each day after install. The shape of the retention curve is one of the strongest predictors of long-tail LTV.
LTV curve: Cumulative ARPU plotted against days since install. The shape of this curve determines payback period and forecast accuracy at each horizon.
Confidence interval: The predicted range around a point forecast at a stated probability (typically 80% or 95%). Wider intervals indicate noisier forecasts.
D-N notation: Shorthand for 'day N after install.' D7 = revenue or retention measured at 7 days; D365 = same metric at one year.
Whale tail: The long, slow-decaying portion of a cohort's LTV curve, driven by a small number of very high-spend users. Strategy, RPG, and 4X games typically have heavy whale tails.
Curve fitting: Fitting a parametric mathematical function (Weibull, exponential decay, power law) to observed cohort data to project forward.
Cross-customer training data: Aggregate cohort data from multiple studios used as a training input, allowing models to learn patterns that no single studio's data would expose.
Drift: The gradual divergence of model predictions from realised outcomes, typically caused by changes in the game economy, audience composition, or attribution environment.

FAQ

Frequently Asked Questions: Cohort Forecasting

Solution

About Kohort's Ktrl Platform

Kohort built Ktrl specifically for the cohort forecasting problem described on this page. Ktrl connects directly to your MMP (AppsFlyer, Singular, or Adjust), trains gradient-boosted prediction models on your studio's own cohort data, and surfaces D90, D180, and D365 ARPU and ROAS forecasts at the campaign level within the first 7 days of each cohort, with explicit confidence intervals at each horizon.

Ktrl is trained on $6B+ of mobile gaming UA spend across studios from hypercasual to mid-core, and is the cohort forecasting platform of choice for studios that want forward LTV signals without building the prediction infrastructure themselves. The same forecasts are used inside leading gaming investors and M&A advisors for diligence on mobile portfolios.

See Ktrl in action

References

Stop hoping your campaigns pay back.

Daily ROAS signals trained on $6bn of UA spend. Free to start, set up in minutes.

Get started free

What is Cohort Forecasting? A Complete Guide for Mobile Game Studios

What you will learn on this page

What is Cohort Forecasting?

Cohort definition

LTV curve fitting

Prediction confidence

Campaign attribution

How Cohort Forecasting Works

Cohort data ingestion

LTV model training

Forecast generation and confidence scoring

ROAS mapping and budget signals

Cohort Forecasting Methods Compared

Common Mistakes in Cohort Forecasting

Aggregating cohorts too broadly

Comparing predicted LTV to incomplete actual LTV

Ignoring confidence intervals when making budget decisions

Conflating retention and revenue curves

Not retraining as the game economy changes

Using cross-genre benchmarks as a forecast input

The Cohort Forecasting Tool Landscape

Mobile Measurement Partners (MMPs)

In-house BI and Data Warehouses

Statistical and ML Stacks

Specialist Cohort Forecasting Platforms

Studios Using Cohort Forecasting: Results

Lotum (4 Pics 1 Word)

52 Entertainment

Apps Teknoloji

Cohort Forecasting Glossary

Frequently Asked Questions: Cohort Forecasting

What is a cohort in mobile gaming UA?

How many installs do you need before cohort forecasting is reliable?

How does cohort forecasting handle seasonality?

Can cohort forecasting work for free-to-play games with very long LTV tails?

How does cohort forecasting handle games with both IAP and IAA revenue?

What is the difference between LTV forecasting and cohort forecasting?

How often do cohort forecasts update?

Can cohort forecasting be used for investor reporting and M&A?

Does cohort forecasting work with all major MMPs?

How does cohort forecasting change with ATT and privacy-first attribution?

Should a studio build cohort forecasting in-house or use a specialist platform?

How accurate are cohort forecasts in practice?

How do confidence intervals work in cohort forecasting?

What are the most common mistakes in cohort forecasting?

About Kohort's Ktrl Platform

Further Reading and External Sources

Stop hoping your campaigns pay back.