forecast

Skill: Forecast

Purpose

Generate time-series forecasts for key metrics using the forecast_helpers library. Supports naive baselines, seasonality detection, and exponential smoothing — enough to answer "what should we expect next?" without complex modeling.

When to Use

• User asks "what will revenue look like next month?" or "forecast DAU"
• After trend analysis reveals a pattern worth projecting
• When sizing an opportunity that depends on future values
• Invoked as /forecast

Invocation

/forecast {metric} — forecast the named metric /forecast {metric} periods=30 — specify forecast horizon /forecast {metric} method=holt_winters — specify method

Instructions

Step 0: Understand the Business Context

Before diving into the forecast, ask clarifying questions if the user hasn't specified:

• What decision depends on this forecast? (e.g., capacity planning, budgeting, staffing, resource allocation)
• Who will use it? (exec summary vs technical deep-dive)
• What's the forecast horizon? (7 days, 30 days, 90 days, a quarter?)
• Are there known upcoming changes? (product launches, campaigns, seasonal events that would invalidate "business as usual" assumptions)

This context shapes how you present results. Capacity planning needs volume impacts and staffing recommendations. Budget planning needs totals and scenario ranges. Executive audiences need decision-focused summaries.

Step 1: Prepare the Data

1. Identify the metric and its source table from the metric dictionary (.knowledge/datasets/{active}/metrics/) or from user specification.
2. Query the data aggregated to the appropriate granularity (daily/weekly/monthly).
3. Create a pandas Series with DatetimeIndex.
4. Clean: forward-fill NaN, drop leading nulls.
5. Validate data sufficiency: Require at least 14 data points for short-term forecasts, 30+ for seasonal forecasts, 60+ for quarterly projections. If insufficient, report: "Not enough history for forecasting — need at least {required} points, have {actual}." Explain what additional data would enable (e.g., "With 30+ days we could detect weekly seasonality").