Forecasting Methodology

Methodology

The forecasting process is based on model comparison and historical validation.

For each data source, multiple candidate models are evaluated to identify the combination of trend, recurring behavior, and short-term dynamics that best explains the observed data and provides the most reliable future predictions.

Candidates are tested against historical observations across multiple forecast horizons and error measures. The selected forecast is therefore based on measured predictive performance rather than a fixed growth assumption or a single forecasting method.

1. Forecasting Objective

The objective is to estimate future usage and cost from the available historical data.

The process focuses on three main questions:

What is the underlying direction of the series?
What recurring or repeating patterns are present?
Which model provides the most accurate predictions for this particular data?

The answer can differ between data sources. Compute, memory, AI usage, storage, and other workloads can have different statistical behavior and are therefore evaluated independently.

2. Temporal Structure

The forecasting process evaluates several types of temporal behavior.

2.1 Trend

A trend represents the longer-term direction of the series.

Candidate models can represent different behaviors, including:

Persistent growth or decline
Changing growth rates
Gradual acceleration or deceleration
A trend that gradually flattens
A recent trend that differs from the longer-term direction
Changes in direction
No meaningful trend

The trend is not selected from a fixed rule. Multiple representations are tested against historical outcomes.

2.2 Periodic Seasonality

Periodic seasonality describes behavior that repeats at a relatively stable interval.

Examples include:

Weekly patterns
Monthly patterns
Yearly patterns
Multiple seasonal patterns occurring together

The presence of a period is established from the historical data. A seasonal component is retained only when it demonstrates predictive value.

2.3 Non-Periodic Recurring Behavior

Some recurring behavior does not follow a fixed calendar period.

Examples can include repeated usage shapes whose timing or duration changes over time.

These patterns are considered separately from fixed-period seasonality. The objective is the same: retain a pattern only when historical validation shows that it improves prediction.

2.4 Combined Structure

Real usage can contain several components simultaneously.

For example:

long-term growth + weekly seasonality + shorter-term fluctuations

The candidate framework therefore evaluates combinations of temporal components rather than forcing the series into a single pattern.

3. Candidate Models

There is no single forecasting model applied to all data.

For each data source, the forecasting process evaluates multiple candidates covering different assumptions about:

Long-term trend
Recent behavior
Seasonal structure
Recurring behavior
Combinations of these components

The candidate set allows the data to determine which representation is most useful.

A candidate is not selected because it provides the closest visual fit to the historical series. Its value is determined by how well it predicts observations that were not available when the forecast was made.

4. Historical Backtesting

Every candidate is evaluated through historical backtesting.

A historical point is treated as a forecast origin. The candidate receives only the information that would have been available at that point and produces a forecast for one or more future horizons.

The forecast is then compared with the actual observations that followed.

This process is repeated over multiple historical forecast origins.

The result is an out-of-sample performance record for each candidate.

This directly measures the property that matters for forecasting: how accurately the model predicts unseen data.

5. Multiple Forecast Horizons

Forecasting performance depends on how far into the future the prediction extends.

A model that is highly accurate for the next few days is not necessarily the best model for several months ahead.

Candidates are therefore evaluated across the forecast horizons relevant to the projection.

This allows the selection process to identify models that preserve useful long-term structure without relying only on short-term accuracy.

6. Multiple Error Measures

Candidate models are compared using multiple forecasting error measures.

Different measures capture different aspects of prediction quality, including:

Absolute size of errors
Error relative to the scale of the series
Sensitivity to large forecasting mistakes

Using several measures reduces dependence on a single definition of forecast quality.

The comparison also considers performance across multiple historical periods and horizons. A candidate must demonstrate useful predictive behavior beyond a single successful period.

7. Model Selection

The final model is selected from the candidates with the strongest and most consistent historical predictive performance.

The process does not assume that the same model should win for every series.

The selected approach can therefore differ according to:

Product or workload
Usage characteristics
Available history
Strength and type of seasonality
Forecast horizon

This is a data-driven selection process. Model choice is based on measured historical performance rather than manual selection for an individual forecast.

8. Long-Term Forecasting

Long-term forecasts rely on persistent patterns rather than simply extending the most recent observation.

The selection process tests whether a candidate can preserve useful long-term direction and recurring behavior over extended historical horizons.

The amount of available history is also important.

Approximately one year of data can provide useful evidence for longer-term direction and many weekly or monthly patterns.

Annual seasonality requires more evidence because one year represents only one annual cycle. Two complete years provide substantially stronger evidence, while three or more years provide stronger support for identifying stable annual behavior.

As the forecast horizon increases, uncertainty also increases. Longer projections therefore require stronger evidence from the historical record.

9. Expected Growth

Where reliable information about expected company or workload growth is available, it can be incorporated into the forecasting process.

This information is treated as an additional input rather than as a replacement for the observed historical behavior.

The objective is to distinguish between:

Growth already visible in the historical data
Expected future growth supported by external information
Short-term fluctuations that should not be extrapolated

10. Reproducibility

The forecasting process is deterministic for a fixed dataset, model configuration, and model state.

The same inputs produce the same candidate forecasts, validation results, and selected model.

This makes the forecast reproducible and allows changes to the methodology to be evaluated against the same historical evidence.

11. Reliability and Limitations

The projection represents the future behavior best supported by the available historical evidence and the tested forecasting candidates.

The strongest evidence comes from patterns that repeatedly demonstrate predictive value during backtesting.

Forecast reliability depends on:

Amount of historical data
Stability of observed patterns
Forecast horizon
Strength of recurring behavior
Changes in the underlying workload

Unexpected changes such as major shifts in usage, product mix, pricing, business activity, or new workloads can cause actual results to differ from the projection.

12. Summary

The methodology can be summarized as:

Collect and prepare historical usage and cost data.
Identify candidate trend and recurring structures.
Build multiple forecasting candidates.
Backtest each candidate across historical forecast origins.
Evaluate performance across relevant forecast horizons.
Compare candidates using multiple error measures.
Select the candidates with the strongest and most consistent predictive performance.
Produce the projection from the selected model.