How AI Predictions Work Inside AllInsights.ai

The Problem with Historical Data Alone

Every game analytics platform can tell you what happened last quarter. Downloads trended up in Brazil, revenue per user dipped in Japan, a competitor's ad spend doubled ahead of their global launch. Historical data is indispensable for understanding where a market has been — but it tells you remarkably little about where a market is going.

Decision makers in gaming face a fundamentally forward-looking set of questions. A UA manager evaluating a new title needs to know whether that game will retain users past day seven before committing budget. A publisher exploring an acquisition needs realistic lifetime-value projections, not just trailing 30-day revenue. An investor reviewing a portfolio company wants to understand whether there is a revenue ceiling for a given genre in a given geography, and how close the title is to hitting it.

Historical averages can approximate answers to these questions, but they break down quickly. Markets shift, player expectations evolve, and the competitive landscape reshapes itself every quarter. A D30 retention benchmark from 18 months ago may no longer reflect the reality of a genre that has since been flooded with new entrants. Cost-per-install figures from last summer say little about what you will pay this winter, when seasonal ad inventory dynamics change the auction entirely.

What teams actually need is forward-looking intelligence — estimates that synthesize the current state of a title, its competitive context, and the broader market trajectory into actionable predictions. That is the gap AllInsights.ai was designed to fill.

Our Multi-Signal Approach

Single-source models are fragile. A model trained only on download counts misses the context that explains why downloads changed. Was it an Apple feature? A viral TikTok campaign? A genre-wide seasonal uplift? Without that context, the model is fitting to noise rather than signal.

AllInsights.ai takes a fundamentally different approach. Our prediction models ingest a broad spectrum of signals, each contributing a different dimension of context:

• App store metadata — descriptions, screenshots, update frequency, pricing changes, and category shifts provide signals about a developer's investment in a title and their strategic direction.
• Ranking trajectories — we track top-chart positions across 140 countries and multiple store categories, capturing momentum that simple download snapshots miss.
• Review sentiment — natural language processing over user reviews surfaces early warnings about technical issues, content staleness, and community reception of new features.
• Ad spend patterns — creative volume, network distribution, and campaign cadence reveal how aggressively a title is being marketed and where it is in its lifecycle.
• Genre benchmarks — every title is evaluated relative to its sub-genre cohort, so our models understand what “good” looks like for a hyper-casual puzzle game versus a mid-core RPG.
• Seasonal patterns — holiday spikes, back-to-school dips, and regional event calendars are encoded as temporal features so models can distinguish seasonal effects from organic growth.
• Competitor behavior — launches, feature updates, and UA pushes from direct competitors create market-level context that affects every title in the category.

By combining these signals into a unified feature space, our models operate on a far richer representation of reality than any single data source could provide. The result is predictions that account for the context behind the numbers, not just the numbers themselves.

What We Predict: 300+ Metrics per Title

Each title in the AllInsights.ai database receives a comprehensive prediction profile that spans the full lifecycle of a game. These are not vague directional scores — they are specific, quantitative estimates that teams can integrate directly into their planning workflows.

Retention predictions

We generate expected retention curves at D1, D7, D30, and D90 horizons. Each estimate is broken down by platform and top geographies. Retention is the single strongest predictor of long-term game health, and getting an early read on expected retention curves lets teams make go/no-go decisions much earlier in the evaluation process.

Lifetime value estimates

LTV projections are computed at 30-day, 90-day, 180-day, and 365-day horizons, segmented by geography and platform. These estimates factor in both in-app purchase revenue and ad monetization, weighted by the title's observed or predicted monetization mix.

CPI benchmarks

Our models estimate realistic cost-per-install ranges by geography and acquisition channel. These benchmarks reflect current market conditions — not static historical averages — because the underlying features update weekly.

Revenue projections with confidence intervals

We model expected revenue trajectories at the title level, including upper and lower bounds. This lets teams understand not just the most likely outcome, but the realistic range of outcomes given current uncertainty.

Engagement and monetization scores

Beyond the core financial metrics, we produce engagement trajectory forecasts, monetization efficiency scores, and market fit indices. These composite metrics help teams quickly assess whether a title is well-positioned within its competitive niche or showing early signs of stagnation.

How the Models Learn

There is no single algorithm that excels at every aspect of game prediction. Tabular features like category rankings, price points, and download volumes are best served by tree-based methods. Sequential patterns — how a game's ranking evolves over weeks and months — require architectures that understand temporal dependencies. And similarity-based predictions (e.g., “titles with this feature profile tend to exhibit this retention pattern”) call for collaborative filtering techniques.

Our production system uses an ensemble approach that combines three model families:

• Gradient-boosted trees (XGBoost / LightGBM) handle the bulk of tabular feature processing. These models are fast, interpretable, and excel at capturing non-linear relationships between features like ad spend, review volume, and revenue.
• Transformer-based sequence models process ranking trajectories and time-series features. By treating a game's daily ranking history as a sequence, these models learn temporal patterns — acceleration, deceleration, plateau signatures — that tree models cannot easily capture.
• Collaborative filtering layers leverage the structure of the game catalog itself. If two titles share similar sub-genre tags, mechanic profiles, and audience demographics, the model can transfer learned patterns from the data-rich title to the data-sparse one.

The outputs of these three model families are combined through a learned meta-model that weights each component based on the prediction target and the available data density for a given title. A newly launched game with limited history will lean more heavily on collaborative filtering, while an established title with months of ranking data will draw more from the sequence model.

Models are retrained on a weekly cadence against the latest data. We use time-aware cross-validation to prevent data leakage, and we calibrate model outputs against publisher-reported metrics where ground truth is available. This calibration step is critical — it ensures that our confidence intervals are well-calibrated, not just that our point estimates are accurate.

Accuracy and Uncertainty

We believe prediction without uncertainty quantification is incomplete. Every prediction AllInsights.ai produces is accompanied by a confidence interval that communicates how much the model “knows” about a given estimate.

Internally, we evaluate model quality using two primary frameworks. The first is MAPE (Mean Absolute Percentage Error), which measures the typical magnitude of prediction errors in percentage terms. For D30 retention predictions — one of the most commonly used metrics on the platform — our models achieve a median MAPE below 15% across the full title catalog. For well-established titles with rich data histories, median errors drop below 10%.

The second framework is calibration. A well-calibrated model should produce 90% confidence intervals that contain the true value roughly 90% of the time. We monitor calibration metrics continuously and retune interval widths when we detect systematic over-confidence or under-confidence in specific segments.

Transparency about limitations is a core principle. Predictions are probabilistic estimates, not guarantees. Several factors can cause real-world outcomes to deviate from predictions:

• Sudden viral events or influencer campaigns that are inherently unpredictable
• Major platform policy changes (e.g., App Tracking Transparency) that reshape market dynamics
• Game updates that dramatically alter the player experience after the prediction was generated
• Macro-economic shifts that change consumer spending patterns across the board

We surface these caveats directly in the product interface and encourage users to treat predictions as one input in a broader decision-making process — not as the sole basis for action.

How to Use Predictions Effectively

Over thousands of conversations with customers, we have developed a set of best practices for getting the most value out of AI predictions:

Use predictions for directional guidance, not exact targets. If our model predicts D30 retention of 8.2%, the actionable insight is not “this game will retain exactly 8.2% of users.” It is that the model expects retention in the range of roughly 7–10%, which positions this title in the middle of its genre cohort. That directional read is enormously valuable for prioritization even if the point estimate is off by a percentage point.

Compare predictions across titles within the same genre. Predictions are most powerful when used for relative comparison. Rather than evaluating a single title's predicted LTV in isolation, compare it to the predicted LTV of the five closest competitors. This benchmarking approach is more robust to systematic model biases because any bias affects all titles in the cohort similarly.

Pay attention to confidence interval width. A narrow confidence interval means the model has strong signal and historical analogues to draw from. A wide interval is a signal that the situation is genuinely uncertain — perhaps the title is in an emerging sub-genre with few precedents, or the game recently pivoted its monetization strategy. Wide intervals are not a failure of the model; they are honest communication about the state of available information.

Combine AI predictions with domain expertise. The best outcomes happen when quantitative predictions and human judgment reinforce each other. If a prediction surprises you, investigate why. Sometimes the model has detected a pattern you overlooked — a competitor's UA ramp-up, a subtle shift in review sentiment. Other times, you have context the model lacks, such as an upcoming content update or a partnership announcement. The most sophisticated users treat AllInsights.ai predictions as a calibration tool for their own intuition, and their intuition as a sanity check on the model.