Usage Forecasting
“What will this account's compute usage be next month?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
What will this account's compute usage be next month?
Usage-based SaaS companies must forecast compute consumption for both infrastructure planning and revenue recognition. A cloud platform serving 2,000 accounts where usage estimates are off by 25% either over-provisions $8M in infrastructure or faces capacity crunches that degrade service for top accounts. Billing surprises from unpredictable usage also drive 18% of customer complaints. The usage trajectory depends on feature adoption, team growth, seasonal patterns, and the account's own business cycles.
Quick answer
Accurate SaaS usage forecasting requires connecting daily usage data, feature adoption events, billing history, and cross-account seasonal patterns in a relational model. The key signal per-account time-series models miss is the usage spike curve that follows feature adoption: accounts that enable batch processing show a 3x usage spike in weeks 2-4 before stabilizing. Graph ML improves forecasting accuracy from 75% to 92%.
Approaches compared
4 ways to solve this problem
1. Per-account time-series (ARIMA/Prophet)
Train a time-series model on each account's historical daily usage to forecast the next 30 days.
Best for
Captures account-specific seasonality and trends. Works well for stable, mature accounts with 12+ months of history.
Watch out for
Each account is modeled independently. When an account adopts a new feature that will spike usage 3x, the time-series model has no way to anticipate it. Also misses cross-account seasonality (Q4 e-commerce spikes).
2. Cohort-based usage projections
Group accounts by industry and plan tier, then project usage based on cohort growth curves.
Best for
Good for capacity planning at the portfolio level. Works when accounts within cohorts behave similarly.
Watch out for
Cohort averages hide individual variance. One account adopting batch processing will spike usage while others in the same cohort stay flat. Over-provisioning for the cohort average wastes infrastructure budget.
3. Committed-tier forecasting
For committed-usage customers, forecast based on their committed tier plus historical overage patterns.
Best for
Simple and works for committed-tier customers where usage is bounded by contract terms.
Watch out for
Does not work for pay-as-you-go accounts, which are the fastest growing and hardest to forecast. Also misses accounts that will exceed their committed tier due to feature adoption or team growth.
4. KumoRFM (relational graph ML)
Connect accounts, daily usage, feature events, billing history, and alert data into a temporal graph. The GNN learns feature-adoption usage curves, cross-account seasonality, and growth trajectory patterns.
Best for
Highest accuracy (92% vs 75% for time-series). Captures the 3x usage spike from new feature adoption and cross-account Q4 seasonality that per-account models miss.
Watch out for
Requires feature-level adoption tracking, not just aggregate usage. If you only log total compute hours without feature attribution, the adoption-curve signal is unavailable.
Key metric: Graph ML improves SaaS usage forecasting accuracy from 75% to 92%, saving $8M in infrastructure over-provisioning and eliminating billing surprises that drive 18% of churn complaints.
Why relational data changes the answer
Usage trajectories are shaped by events in other tables: feature adoption (an account enabling batch processing will spike compute 3x), team growth (5+ new users per month accelerates API calls), and cross-account patterns (all e-commerce accounts spike in Q4). A per-account time-series model with 'last_month_hours = 3,400' projects a smooth growth curve and misses that this account just adopted batch processing two days ago, which means a 3x spike is coming in weeks 2-4.
Relational models connect usage data to feature adoption events and cross-account seasonality patterns. They learn that accounts adopting batch processing follow a specific usage curve (3x spike, then stabilization at 1.5x baseline) and that similar-industry accounts share seasonal patterns. For a platform serving 2,000 accounts, improving accuracy from 75% to 92% saves $8M in infrastructure over-provisioning. It also eliminates billing surprises: when the model predicts a usage spike, the account team can proactively notify the customer rather than surprising them with a bill 25% higher than expected.
Forecasting account usage from historical trends alone is like forecasting household electricity from last year's bills. You miss that the family just bought an electric vehicle (feature adoption), that they are having a pool installed (team growth), and that it is going to be the hottest summer on record (cross-account seasonality). The connected context around the usage number matters more than the number itself.
How KumoRFM solves this
Graph-learned product intelligence across your entire account base
Kumo connects accounts, daily usage, feature events, billing history, and alert patterns into a temporal graph. It learns that accounts that recently adopted the batch-processing feature show a 3x usage spike in weeks 2-4 before stabilizing. The model captures seasonality (e-commerce accounts spike in Q4), growth trajectories (accounts adding 5+ users per month accelerate usage), and cross-account infrastructure patterns (when a shared cluster's accounts all grow simultaneously, capacity planning must account for contention).
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
ACCOUNTS
| account_id | plan | industry | active_users | contract_type |
|---|---|---|---|---|
| ACC401 | Usage-based | E-commerce | 85 | Pay-as-you-go |
| ACC402 | Usage-based | Finance | 30 | Committed |
| ACC403 | Usage-based | Technology | 120 | Pay-as-you-go |
USAGE_DAILY
| usage_id | account_id | date | compute_hours | storage_gb | api_calls |
|---|---|---|---|---|---|
| UD01 | ACC401 | 2025-03-01 | 450 | 2800 | 125,000 |
| UD02 | ACC401 | 2025-03-02 | 480 | 2850 | 132,000 |
| UD03 | ACC402 | 2025-03-01 | 120 | 500 | 28,000 |
FEATURES
| feature_id | account_id | feature | enabled_date | usage_30d |
|---|---|---|---|---|
| FT01 | ACC401 | Batch processing | 2025-02-15 | High |
| FT02 | ACC401 | Real-time API | 2024-12-01 | High |
| FT03 | ACC402 | Batch processing | 2025-03-01 | New |
BILLING
| billing_id | account_id | month | amount | vs_estimate |
|---|---|---|---|---|
| BL401 | ACC401 | 2025-02 | $12,400 | +18% |
| BL402 | ACC401 | 2025-01 | $10,200 | +8% |
| BL403 | ACC402 | 2025-02 | $3,100 | -5% |
ALERTS
| alert_id | account_id | type | timestamp | threshold_pct |
|---|---|---|---|---|
| AL01 | ACC401 | Usage spike | 2025-02-20 | 150% |
| AL02 | ACC403 | Quota warning | 2025-03-01 | 90% |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(USAGE_DAILY.COMPUTE_HOURS, 0, 30, days) FOR EACH ACCOUNTS.ACCOUNT_ID
Prediction output
Every entity gets a score, updated continuously
| ACCOUNT_ID | INDUSTRY | LAST_MONTH_HOURS | PREDICTED_NEXT_MONTH |
|---|---|---|---|
| ACC401 | E-commerce | 12,800 | 16,200 |
| ACC402 | Finance | 3,400 | 5,800 |
| ACC403 | Technology | 8,900 | 9,100 |
Understand why
Every prediction includes feature attributions — no black boxes
Account ACC402 -- Finance, 30 active users
Predicted: 5,800 compute hours predicted next month (+71%)
Top contributing features
New batch processing adoption
Enabled 2 days ago
32% attribution
Similar-account batch adoption curve
3x spike, weeks 2-4
24% attribution
User growth rate (30d)
+5 new users
18% attribution
Historical monthly growth
+12% MoM avg
14% attribution
Committed tier buffer
80% of committed used
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about usage forecasting
How do you forecast SaaS compute usage per account?
Connect daily usage data, feature adoption events, team growth metrics, and billing history into a relational model. The model learns account-specific growth trajectories, feature-adoption usage curves, and cross-account seasonal patterns. Graph ML forecasts 30-day usage with 92% accuracy vs 75% for per-account time-series models.
Why are per-account time-series models inaccurate?
Per-account models treat each account as an island. They cannot anticipate usage changes driven by feature adoption (batch processing causes a 3x spike), team growth (new users adding API calls), or cross-account seasonality (all e-commerce accounts spike in Q4). These signals live in related tables, not in the account's own usage history.
How does feature adoption affect usage forecasting?
Feature adoption is the single biggest source of forecasting error. An account enabling batch processing shows a 3x compute spike in weeks 2-4 before stabilizing at 1.5x baseline. Per-account models miss this entirely. Relational models learn the specific usage curve associated with each feature adoption and apply it proactively when an account enables a new feature.
How do you prevent billing surprises in usage-based SaaS?
Billing surprises drive 18% of churn-related complaints. Accurate usage forecasting lets account teams proactively notify customers when a usage spike is predicted. This turns a negative billing surprise into a positive customer conversation and an opportunity to discuss committed-tier upgrades that save the customer money.
What is the ROI of better usage forecasting in SaaS?
A platform serving 2,000 accounts that improves forecasting accuracy from 75% to 92% saves $8M in infrastructure over-provisioning annually. It also reduces churn from billing complaints by eliminating usage surprises. The combined impact of infrastructure savings and retained revenue makes usage forecasting one of the highest-ROI ML applications in SaaS.
Bottom line: A usage-based SaaS platform serving 2,000 accounts that improves usage forecasting accuracy from 75% to 92% saves $8M in infrastructure over-provisioning and eliminates billing surprises that drive 18% of churn-related complaints. Kumo captures feature-adoption usage curves and cross-account seasonality that time-series models on individual accounts cannot learn.
Related use cases
Explore more B2B SaaS use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
