Plonky Forecasting Skill

You are using Plonky, a time-series forecasting API. This skill teaches you the end-to-end forecasting workflow — from data preparation to interpreting results.

• MCP server setup & source: https://github.com/addysmoke/plonkyai_mcp
• Full API docs: https://plonky.ai/agents

Tools Available

You have access to these MCP tools:

• register — create an account and get an API key (only if not already authenticated)
• get_credits — check your credit balance
• upload_data — upload CSV data
• list_datasets — see uploaded datasets
• analyze_dataset — get summary stats and data quality info
• create_forecast — run a forecast (blocks until complete)
• get_forecast — retrieve results for an existing forecast
• create_backtest — evaluate forecast accuracy (period_count + period_type, not n_splits)
• create_forecast_batch — forecast across multiple dimensions

When to Use Forecasting

Before reaching for Plonky, assess whether a statistical forecast is the right tool for the task.

Good fit:

• Recurring time-series data with enough history (revenue, traffic, demand, usage metrics)
• Planning and budgeting horizons (next quarter, next year)
• Capacity and inventory planning where directional estimates save money
• Comparing segments or scenarios ("which region is growing fastest?")

Poor fit — tell the user why and suggest alternatives:

• Brand-new products or markets with <2 months of data. No model can extrapolate from nothing — suggest manual estimates or analogous-product benchmarking instead.
• One-off or event-driven outcomes (will this product launch succeed? what will Q4 revenue be if we change pricing?). These are decisions, not time-series patterns.
• Highly irregular or random data where past patterns have no predictive value (e.g., individual customer churn dates, lottery outcomes). If the data looks like noise, say so.
• Real-time or sub-hourly granularity. Plonky is designed for daily-granularity series.
• Weekly or monthly data can be used but results will be less reliable. Plonky internally converts all data to a daily time series (filling gaps with zeros or forward-fill). With monthly input, this creates ~30 artificial data points per real observation, which degrades forecast quality. If the user only has monthly data, warn them that results are directional estimates at best.
• Data dominated by external decisions (marketing spend, pricing changes, policy shifts). A univariate forecast won't capture interventions — the user needs causal modeling or scenario analysis, not extrapolation.

When in doubt, run the forecast and a backtest. If MAPE is >30%, the data may not be forecastable — tell the user honestly.

Workflow

Follow these steps in order. Do not skip the analysis step.

Step 0: Check Authentication