Quick Start
Getting Started
Load the bundled demo data or upload a CSV to begin. The platform supports wide-format tables with one date column, optional dimensions, and one numeric column per metric.
Tip
Start with the built-in demo data to explore the platform without login. Connect only when uploading your own CSV.
1. Load Data — Load demo data or upload a CSV file.
2. Model Fitting — Configure parameters, train candidate models, backtest with holdout periods, and inspect leaderboard diagnostics.
3. Forecasting — Review BOY planning forecasts, residual-scaled prediction intervals, and run periodic same-method reforecasts.
Key Concepts
Wide Format
A table where each metric is a column: date, channel, persona, weekly_active_users, new_trials. This is the default workflow.
Metrics and Dimensions
A metric is the value to forecast, such as revenue or active users. A dimension is a way to break it down, such as region, channel, or customer segment.
Fitting Levels
Aggregate: Fit the total directly. Segments Rollup: Fit each segment independently, then sum to produce the total forecast.
Backtesting
Holdout the most recent N periods, fit on the remaining history, then measure forecast accuracy on the withheld actuals.
Horizon and Holdout
The horizon is how far ahead the forecast runs. The holdout is recent historical data set aside to test a candidate method before it is used for planning.
Method Selection
The platform can compare several candidate methods using the same holdout data. Accuracy, bias, interval coverage, and business context help determine which method is most suitable for the forecast.
Governed Run Metadata
API-backed forecasts include data-quality status, selected method version, parameter hash, lineage IDs, draft approval state, and backtest artifact references.
Prediction Intervals
Use holdout residual scale and horizon growth to produce clear prediction intervals around the fitted model. In the annual-planning demo, the upper edge of a central 60% interval is treated as a one-sided P80 planning scenario under the model's symmetric-residual assumption.
Snapshots and Reforecasts
A snapshot preserves an approved planning forecast. A reforecast reruns the forecast as new actuals arrive, so teams can compare the current outlook with the original plan.
API Reference
Base URL & Auth
Production: https://forecasting-agent-platform-154742125932.us-central1.run.app/api. Local development uses http://127.0.0.1:8000 behind the Vite proxy. Authenticated endpoints require Authorization: Bearer <token>; public demo endpoints use X-Demo-Session and are limited to bundled demo data.
Data & Session Endpoints
POST /auth/frictionless-login issues a user-scoped bearer token. POST /metrics/upload-csv-text registers custom CSV data for an authenticated tenant. POST /demo/metrics/upload-csv-text registers bundled demo data without login. GET /datasets/latest restores the latest authenticated dataset.
Forecast Execution
POST /forecasts/run accepts metric_id, segment_id, method_id, frequency, horizon_periods, holdout_periods, interval_width, selection_metric, optional training_start_date/training_end_date, and model parameters such as season length, holidays, covariates, lag windows, rolling-origin settings, and reconciliation. POST /demo/forecasts/run runs the same contract for public demo sessions.
Snapshots, Reports & Agent
POST /forecasts/snapshot saves a governed forecast artifact. GET /forecasts/snapshots lists saved artifacts. POST /forecasts/compare-snapshots compares original and refreshed plans. POST /ai/e2e-report creates a grounded report. POST /agent/run powers Workbench and Agent Conversation flows so the LLM can plan, critique, narrate, and orchestrate while Python owns the math.
Choose the Right Connection
Use the REST API when an application, scheduled job, or data workflow needs a predictable request and response. Use MCP when an AI assistant needs to discover the available forecasting tools, choose the next action, and explain the result in conversation.
Private MCP Pilot
How It Connects
The pilot uses the standard MCP Streamable HTTP transport at /mcp/. Approved participants receive the connection details and use their existing tenant access; the service remains private and is not a public MCP endpoint.
Available Tools
forecasting.list_metrics, forecasting.describe_metric, forecasting.actual_summary, forecasting.run_forecast, forecasting.create_snapshot, forecasting.list_snapshots, and forecasting.compare_snapshots. Upload data through the web app or REST API, then use MCP to inspect and forecast approved tenant data.
How the Pieces Fit Together
API: a direct request from one application to another. MCP: a standard way for an AI client to discover and call approved tools. Skill: reusable instructions that guide an agent through a task. Plugin: a package that distributes skills and connection setup. In this pilot, skills guide the workflow, the plugin distributes it, and MCP performs the approved tool calls.
Access and Data Boundaries
MCP does not give an agent unrestricted access to the platform. Each call is authenticated, limited to the participant's tenant data, and performed by the governed backend. The private pilot keeps the connection and its tool set limited while the workflow is evaluated.
Recommended Workflow
Start by listing metrics and describing the chosen one. Review actuals, run one scoped forecast, then save and compare snapshots when you need a planning record. The backend owns the forecast math, model scoring, snapshot immutability, and tenant boundaries; the client turns those results into an analyst-friendly conversation.