Forecasting mobile game profitability: How publishers predict UA performance

• Early user acquisition performance - as early as day 7 - can reliably predict long-term mobile game profitability.
• Small changes in early monetisation signals can shift annual ROAS projections and UA budget allocation.

Iaroslav Kobozev is head of marketing analytics at AppQuantum.

Profit forecasting is a continuous process that affects both a studio’s budget and the way user acquisition is managed. It is especially critical for early-stage studios that do not yet have a strong financial buffer. In our work, we regularly update a system of coefficients so that decisions about launching, pausing, and scaling advertising campaigns are made with controlled risk.

The core logic is simple: we project annual ROAS (D365) from actual D7 performance. One week of data is already sufficient to produce a relevant and reasonably valid forecast. At the project–platform level, this projection typically stays with 5–10% margin of error. A one-year horizon is where budgeting and P&L ultimately align.

The forecast directly affects cash flow and risk exposure and determines how budgets are distributed across projects, advertising sources, GEO splits, platform splits, and optimisation strategies.

It forms the basis for early decisions on pivoting or scaling a project, influences revenue planning and P&L, and shapes content update schedules and live ops planning. Both overprediction and underprediction are harmful scenarios: in one case, capital is overspent; in the other, investment is held back and potential upside is lost.

Data passes through a full analytical pipeline: collection →  cleaning → normalisation into a unified logic → aggregation → naming standardisation → statistical transformation for forecasting → mathematical modelling → reverse transformation → and finally decomposition into the required aggregations.

For a reasonably accurate forecast, a single cohort should contain no fewer than 300 installs or 10 paying users. This is a baseline minimum, though in some cases it becomes necessary to lift aggregation upward by optimisation type → source → country, or similar dimensions in order to reduce variance.

This applies equally to new projects coming in for evaluation. Quite often, developers fail to segment data properly, drawing incorrect conclusions about a project’s profit potential and creating operational problems down the line. The most frustrating part is that these issues are easily avoidable if proper tagging and aggregation are implemented from the start.

If historical data turns out to be heavily contaminated, it is sometimes possible to work with shorter time windows and industry benchmarks, but that is an entirely different story.

To enable even basic analytics, an SDK must be embedded in the game, core product events must be properly tagged, and payment integrations configured. The resulting data then needs to be stored and structured. With this foundation in place, it becomes possible to run simple end-to-end analytics and build basic forecasts.

This includes funnel analysis, retention tracking, average purchase size, daily revenue dynamics, daily cost dynamics, and a baseline projection of how much revenue the game will generate over a month, a quarter, or a year at the product level.

From there, analytics progresses to a more advanced stage. This includes identifying player clusters and segments, uncovering behavioral patterns by GEO or acquisition source, and analysing how different cohorts move through content. Advanced forecasting then answers questions such as how much revenue a specific optimisation strategy will generate on a given traffic source, for example on Facebook.

Why D7 became sufficient for operational forecasting

A one-week window fits naturally into UA operations for two reasons. Over this period, enough events and payments accumulate to support decisions without excessive delay, although in some cases creatives are allowed to run slightly longer to stabilise.

At the same time, we do not force all projections to rely on D7 regardless of statistical volume. If data density is insufficient at a granular level, the forecast is built at a higher aggregation level where variance is lower. This depends on the type of game and on what share of total lifetime revenue is realised early. Typically, at least 10–20% of final revenue needs to materialise in early stages for the projection to be considered valid.

The D7 to D365 linkage creates a predictable budget management rhythm. Each week, actual D7 performance is measured, the annual forecast is updated, and clear operational decisions follow: whether to maintain spend, accelerate, slow down, reallocate across sources, or restart creative packages. Additional calibration checks are performed at D30, D90, and D180 to prevent systematic drift toward the end of the year.

We also review forecast errors on a weekly basis and incorporate corrections into the models. This is critical because the advertising market is dynamic, data distributions evolve, and forecasting approaches must evolve alongside them. 

For this reason, our models are never static.

It is worth briefly touching on composite metrics. Relying on a single coefficient introduces instability, so in practice we use a weighted combination of multiple signals. The anchor is actual D7 ROAS, which is supplemented by payer share at D3 and D7, early revenue per install at D1 and D3, session frequency, and the proportion of high-value purchases among the top paying percentiles. Together, these signals capture both the speed of early monetisation and the potential of the long-tail revenue curve.

The weights are not universal. For products with early revenue peaks, early signals carry more influence, while for midcore games with long LTV curves, later signals dominate. For example, 4X and hard-midcore titles place greater weight on late-stage factors and high-ticket purchase share, while match-3 and idle games emphasise early ROAS, retention, and payment base stability.

New projects begin with the closest template and gradually retrain weights on their own cohorts. Ablation checks are run regularly to assess how accuracy changes when individual signals are removed, keeping the composite metric lean and free of unnecessary noise. Composite modelling deserves a separate deep dive, so we will move on.

Coefficient adaptation

Our coefficients drift alongside changes in the product, traffic mix, and seasonality. The model itself updates daily, while methodological improvements are introduced several times per month following systematic error analysis as it was mentioned earlier.

Coefficients are recalculated weekly on a rolling one-year window, with the most recent three to five days excluded due to attribution delays. Typical adjustments range around 0.1–0.2% per coefficient, which is sufficient to reflect new data without introducing volatility.

Before any refit, data quality is thoroughly validated. Cohort integrity is checked with install date as the starting point, currencies and time zones are verified, spend is reconciled between MMPs and ad platforms, and attribution windows are synchronised.

Any anomaly - such as install spikes without corresponding clicks, abnormal payer shares, or spend discrepancies - blocks coefficient updates and triggers an investigation. Updating models on corrupted data is a direct path to increased forecasting error.

Drift is monitored continuously. If distributions of key features or average error exceed defined thresholds, an unscheduled review is triggered.

During major product changes involving economy, pricing, or attribution logic, routine updates are frozen and a separate testing loop is activated.

Rollouts after refits proceed in stages, starting with offline backtesting, followed by limited deployment across selected sources or GEOs, and only then full scaling. If the error exceeds acceptable limits at any stage, automated rollback is triggered.

Forecast accuracy

Accuracy is calculated on closed annual cohort slices and aggregated using spend-weighted averages. For 2023 cohorts, the average error was 4%. For cohorts from July 2023 through July 2024, it was 5%. These figures reflect weighting by real budgets rather than simple averages across projects.

Mature products consistently remain within ±5% without systematic bias. New or rapidly scaling projects are more challenging; underestimation of up to 10% occurs regularly. The practical reason is that payer behaviour shifts faster than the model can fully adapt. We track this error segment separately, and when systematic underestimation appears within growth or uplift segments, corrective multipliers are applied and weight recalibration is accelerated.

Segment-level accuracy is mandatory. Separate error profiles are maintained across platforms, GEOs, traffic sources, and optimisation types. If a slice lacks sufficient statistical volume, the forecast is lifted to a higher aggregation level. Reports clearly flag zones where scaling is not recommended until sufficient data accumulates. These are not cosmetic indicators but concrete budget management rules.

We also analyse performance across multiple horizons - D30, D90, D180, and D365. If D7 alignment is strong but the long-term margin of error grows, the contribution of late-stage factors within the composite is increased. If the early margin of error is higher but convergence improves by D365, the model is allowed more flexibility in the early window to capture creative and bid effects faster.

Practical stop signals that protect budget

• A sharp increase in payer share without a corresponding rise in revenue per payer - triggers tracking and anti-fraud validation, and coefficient updates are frozen until verification is complete.

• Spend discrepancies between MMP data and ad platform dashboards beyond acceptable thresholds - recalculation is blocked and exports are reconciled.

• Systematic underestimation in growth and uplift segments - corrective multipliers are applied and weight recalibration for those segments is accelerated.

Key figures and thresholds we rely on:

• Forecast error across closed 2023 annual cohorts: 4%

• Forecast error across cohorts from July 2023 to July 2024: 5%

• Mature products remain within ±5% margin of error

• New or rapidly growing products may show up to 10% underestimation

• Minimum viable cohort volume: 300 installs or 10 payers per week

• Monthly coefficient adjustments: roughly 0.1–0.2% per coefficient, affecting around 60% of the set

All of the above keeps forecasting error within ranges that allow confident financial control: overall 4–5% on annual horizons for mature products and controlled deviations up to 10% for new or fast-scaling ones.

In practical terms, this means scaling or shutdown decisions are not driven by intuition, but by defined variance ranges and clear operational thresholds.

Ultimately, an effective forecasting system isn’t built around a single “secret coefficient,” but around data discipline, composite signal design, continuous adaptation, and rigorous testing that only promotes truly strong models into production. When these components align, forecasting becomes a reliable foundation for growth rather than a source of surprises.