forecasting-timeseries

Forecasting Time Series

Quick Start

1. Query time series data with the lakehouse (daily granularity, 30-90 days covers both tools well)
2. Use Analyze Time Series Insight to detect anomalies and get a statistical forecast
3. If you need higher accuracy or uncertainty bands, follow up with Forecast with Chronos (needs 14+ days, best with 30+)

When to Use This Skill

• User asks about trends, spikes, or drops in a metric over time
• User wants to predict or forecast future values using local analysis tools
• User asks "is this normal?" about a metric value
• Investigating anomalies or unexpected changes in an ad-hoc session
• User asks for projections, predictions, or what to expect next week/month
• Keywords: "forecast", "predict", "anomaly", "spike", "drop", "trend", "projection", "unusual", "normal range"

This skill vs. forecast tasks: This skill runs on-demand, local analysis using Python tools (Chronos, statsmodels). For recurring, automated forecasting on a schedule, use the configuring-tasks skill to create a forecast task instead.

Two Tools, Two Purposes

There are two complementary time series tools. Use one or both depending on the question.

	Analyze Time Series Insight	Forecast with Chronos
Best for	"What happened?"	"What will happen?"
Anomaly detection	Yes (Z-score + IQR)	No
Forecasting method	Exponential smoothing	Chronos-2 ML model
Uncertainty bands	No	Yes (10th/90th percentile)
Minimum data	3 days	14 days
Recommended data	14-90 days	30-365 days
Default horizon	7 days	14 days
Max input size	2000 chars	3000 chars

When to Use Each

User question about a metric over time
  │
  ├─ "Is this value normal?" / "Why did X spike?"
  │   → Analyze Time Series Insight (anomaly detection)
  │
  ├─ "What will happen next week?" / "Forecast revenue"
  │   → Forecast with Chronos (ML forecast with uncertainty)
  │
  └─ "Analyze this trend and predict what's next"
      → Both: Analyze first, then Chronos for deeper forecast

Workflow

Step 1: Query the Data

Use query_lakehouse to get daily time series data. Format the result as:

[{"date": "2024-01-01", "value": 100}, {"date": "2024-01-02", "value": 105}]

Keep the data compact. Aggregate to weekly if the date range exceeds 90 days. The tools have strict character limits on input.

Step 2: Run Analysis

Start with Analyze Time Series Insight for a statistical overview:

• Detects anomalous values using Z-score and IQR methods
• Identifies whether the latest value is anomalous
• Provides trend direction and day-over-day change
• Generates a short-term exponential smoothing forecast

Step 3: Enhance with Chronos (Optional)

If the user needs a more accurate forecast or wants confidence intervals, run Forecast with Chronos:

• Provides point forecast plus 10th/90th percentile uncertainty bands
• Better at capturing complex seasonal patterns
• Indicates forecast confidence (high/medium/low based on band width)

Interpreting Results

Anomaly Detection

The analysis tool flags anomalies using two combined methods:

• Z-score: Values more than 2 standard deviations from the mean
• IQR: Values below Q1 - 1.5 * IQR or above Q3 + 1.5 * IQR

A value flagged by either method is reported. If the latest value is anomalous, it requires attention.

Trend Direction

Forecast Change	Interpretation
> +5%	Increasing trend
-5% to +5%	Stable
< -5%	Decreasing trend

Chronos Confidence Levels

Uncertainty Band Width	Confidence
< 20% of forecast mean	High — narrow band, reliable
20-50%	Medium
> 50%	Low — wide band, treat with caution

Common Pitfalls

1. Sending too much data — Tools have 2000/3000 char limits. Aggregate to weekly for long ranges
2. Too few data points — Chronos needs 14+ days. Analysis needs 3+ but works best with 14+
3. Using Chronos for anomaly detection — Chronos only forecasts; use Analyze Time Series for anomalies
4. Skipping aggregation — Hourly data quickly exceeds size limits. Always use daily or weekly granularity
5. Ignoring uncertainty bands — A Chronos forecast with wide bands means low confidence; communicate this clearly
6. Not checking seasonality — Weekly patterns (weekday vs weekend) need at least 14 days to detect
7. Forecasting without context — Always pair forecasts with what the current trend shows

References

• Chronos forecasting details - Read when generating forecasts or configuring prediction horizons and confidence intervals
• Anomaly detection methods - Read when detecting anomalies or choosing between statistical methods