Welcome to Forecast Coach Documentation
Forecast Coach brings state-of-the-art time-series foundation models into your daily AI discussions. Here is everything you need to get the most out of your forecasts.
Spreadsheet Formatting Best Practices
Forecast Coach automatically extracts time series from Excel and CSV files. For optimal results, ensure your spreadsheet has:
- Date/Timestamp Column: Regular ascending intervals (daily, weekly, or monthly dates like
2024-01-01). - Observed Value Column: Clean numbers (e.g. units sold, revenue, pageviews). Missing observations are handled seamlessly.
- Multiple Series / SKUs: If your sheet has multiple columns (e.g. Sales_Region_North, Sales_Region_South), you can ask the AI to forecast any specific column or iterate through them.
Prompt Templates to Try
“Forecast the next 8 weeks of demand from this sheet and display the prediction chart with P10-P90 bounds.”
“What does the forecast look like if we see a 15% promotional bump in week 4?”
“Backtest the forecast on the last 3 months of historical data to evaluate how accurate the model would have been.”
Excel Demo Pack & Educational Scenarios
To help you evaluate Forecast Coach immediately, we provide an all-in-one demonstration Excel workbook containing 6 synthetic scenarios, realistic business patterns, and tested prompts.
Download Forecast_Coach_Demo.xlsx
8 worksheets • 6 business scenarios • Ready-to-copy prompts • Zero configuration
The 6 Synthetic Scenarios
| Worksheet | Freq | Rows | Horizon | Demonstrates |
|---|---|---|---|---|
| Daily_Retail | D | 182 | 14 days | Start here: gradual trend growth, weekend seasonality, realistic variance. |
| Stockouts_Zeros | D | 182 | 14 days | Missing data: 7 stockouts stay blank (null), while 2 zero-sale days stay numeric 0. |
| Open_MWF | D | 182 | 14 days | Sparse operating calendar: stall open Mon/Wed/Fri (105 intentional blank closures). |
| Weekly_Orders | W | 104 | 8 weeks | Sunday week-ending labels spaced exactly 7 days apart; 2-year trend and annual cycle. |
| Monthly_Revenue | MS | 60 | 6 months | 5 years of SaaS revenue with month-start (MS) dates; Q4 holiday peaks. |
| Multi_SKU | D | 182 | 14 days | 3 products on 1 calendar (Blue: growing, Red: stable, Bundle: declining). |
Input Contract & Guardrails
- Actual Blanks vs. Numeric Zeros: Empty value cells serialize as
null. Never delete dates or fill out-of-stock days with 0. A numeric0signifies verified zero demand. - One Regular Grid Per Call: Pass rows in ascending chronological order without skipping dates. Daily includes weekends; weekly is exactly 7 days apart; monthly uses month starts (
MS). - Target Selection Guardrail: When multiple target columns exist (such as in
Multi_SKU), always specify the target column. If no target is specified, the assistant should ask for clarification rather than arbitrarily guessing. - Cadence Mismatch Guardrail: Requesting a daily forecast on weekly dates is a deliberate mismatch. The assistant will decline and warn of the cadence mismatch rather than silently corrupting or resampling dates.
- Empirical Coverage Basis: Always report the tool's empirical
band.coverageand measurementband.basis(e.g. 80.3% measured on historical e-commerce for Chronos) rather than assuming a nominal 80%.
Supported Models and Ensembles
Forecast Coach provides a comprehensive multi-model suite running in-process: Amazon Chronos-Bolt for zero-shot deep learning, Meta Prophet for strong calendar and holiday seasonality, NeuralProphet for hybrid deep autoregression, and Automated Median Consensus Ensembles. The built-in rolling-origin backtesting engine automatically evaluates MASE errors against Seasonal Naive 7 to recommend the winning algorithm for your data.
Need Help or Enterprise Setup?
For private VPC or on-premises deployment connecting directly to SAP, Oracle, or PostgreSQL, email us at guy@mlguy.us.