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").

Step 2: Detect Seasonality

Run detect_seasonality() from helpers/stats/forecast_helpers.py:

• If seasonality detected, report: "Found {strength} {period}-day seasonality."
• Store the dominant period for use in Step 3.

Step 3: Generate Forecasts

Run multiple methods and compare:

1. Naive (last value): naive_forecast(series, periods, method='last')
2. Naive (seasonal): If seasonality detected: naive_forecast(series, periods, method='seasonal_naive')
3. Exponential smoothing (auto): exponential_smoothing(series)
4. Holt-Winters: If seasonality detected and enough data: exponential_smoothing(series, seasonal_period=dominant_period)

Compare MSE across methods. Select the best-fit method.

Step 4: Generate Chart

Using chart_helpers:

1. Call swd_style()
2. Plot historical data as a solid line
3. Plot forecast as a dashed line with lighter alpha
4. Add confidence band (±1 std of residuals) as shaded area
5. Mark the historical/forecast boundary with a vertical dashed line
6. Use action_title() with a forward-looking title
7. Save to working/forecast_{metric}_{DATE}.png using save_chart()

Step 5: Present Results

Structure the output based on business context from Step 0:

Always include:

• Data validation summary: "{N} days of historical data analyzed, sufficient for {horizon}-day forecast"
• Best method and why (lowest MSE on validation holdout)
• Seasonality findings (if detected)
• Forecast values for key milestones (next 7/14/30/60/90 days as relevant)
• Confidence intervals (±1σ and ±2σ)
• Standard caveats: "Forecasts assume past patterns continue. External factors not modeled."

Business context translation:

• Capacity planning: Translate forecast to volume impacts (e.g., "+4% conversion → +4% order volume → scale fulfillment by 4%"). Provide staffing/resource recommendations with risk buffers.
• Budget planning: Provide period totals (monthly, quarterly) with upside/downside scenarios. Include sensitivity to key assumptions.
• Monitoring/alerting: Suggest thresholds for when actuals diverge from forecast (e.g., "If Week 1 revenue < $X, revisit assumptions").
• Exec summary: Lead with the decision implication, then support with numbers. Example: "No team scaling needed — conversion increase (+4% over 2 weeks) is within current capacity."

The forecast numbers are necessary but not sufficient. Your job is to answer "so what?" for the user's planning context.

Rules

1. Always run at least 2 methods for comparison. The user needs to know the forecast isn't arbitrary — showing that one method outperforms others builds confidence.
2. Never present a forecast without stating assumptions. Every forecast has an expiration date and boundary conditions. Make them explicit.
3. Always include a naive baseline. Users need to see if the model adds value beyond "just use the last value." If naive performs nearly as well, say so — simpler is better.
4. Flag if residuals show systematic patterns. Autocorrelated residuals mean the model is missing something (trend, seasonality, structural break). Don't paper over this — it affects forecast reliability.
5. If the data has a structural break, warn that forecasts may be unreliable. A sudden spike or drop in the last N days means the historical pattern may not continue. Surface this uncertainty.
6. Translate to business impact automatically. Don't just report "4.09% conversion on Day 14" — connect it to the planning question: "4.09% conversion (+4% vs today) means approximately +4% order volume, which is within your current fulfillment capacity." The forecast is a means to a decision, not the end goal.

Edge Cases

• Constant series: Report "No variation — forecast is the constant value"
• Strong trend + no seasonality: Use Holt's (double) exponential smoothing
• Very short history (<30 points): Only use naive methods, warn about accuracy
• Data gaps: Interpolate or warn, depending on gap size