MCP App Store
by KeenEthics
Forecast Coach icon

Forecast Coach MCP Server for Claude

By MLGuy Inc.

Overview

Your conversational time-series forecasting partner. Drop Excel workbooks or CSV files into your everyday AI chat client and get instant, foundation-model forecasts with confidence bands and interactive charts.

How to connect

  1. Claude: Open the Forecast Coach listing in the Claude connectors directory and choose Connect. Claude listing Forecast Coach exposes 2 tools here.

Tools

backtest

Claude
Hold out the last horizon periods across origins rolling holdout windows and evaluate registered forecasting algorithms (Chronos, NeuralProphet, Prophet, seasonal naive, moving average, seasonal median) and the median ensemble. Returns MASE scores against the seasonal-naive floor, ranking models side by side. Declared regressors and holidays are fitted on prophet only, which is then scored with and without them on the same origins.

forecast_series

Claude
Forecast intake policy Forecast a daily, weekly or monthly business time series and return the forecast, a confidence band, and the measured coverage of that band, drawn as a chart panel beside your reply. Horizon: 1 to 64 periods in one call. A longer horizon is refused — ask again from the end of the first forecast rather than asking for a year in one go. Title: title names what the series is, in the operator's own words — "Daily covers, Harbour Street". It becomes the panel heading, the line the operator reads above their own chart. Ask which column — never pick one If the sheet has more than one candidate value column — several SKUs, several channels, several stores — do NOT pick one. Ask which the operator wants. Taking the leftmost column is a positional default, not a semantic one. An operator who skims the caveat silently receives a forecast of the wrong product. State the options and stop: > This sheet has three value columns — SKU-1041 Blue, SKU-1042 Red and SKU-2200 Bundle. > Which should I forecast, or would you like the combined total? Three cases, not two: when a day is null, when it is 0, and when it is closed A day nobody could observe, a day that genuinely sold none, and a day the business was shut are three different facts. The y array tells the first two apart; the third is a fact about the calendar and is declared as one. 1. Stockout, unrecorded, or not yet measured — send `null`. Demand existed and could not be served, or could not be seen. A stockout fed as 0 teaches the model a slump that never happened, and the forecast then under-orders into the next one. Look for the evidence: an adjacent stock or availability column, a run of consecutive zeros in an otherwise active series, or a weekday pattern. If the sheet gives you no way to tell a stockout from a genuinely quiet day, say so rather than choosing silently. 2. Open, and genuinely sold none — send `0`. A real observation of zero, and it belongs in the series as one. 3. Closed, or not trading — declare it in the calendar. Put a standing weekly closure in closed_weekdays (["sat", "sun"], or the full English names) and any one-off closure, past or future, in closed_dates (["2026-12-25"]). Send the row itself as 0 or null, either one: once a day is declared closed, its row is fitted as unobserved whichever it carries, and every forecast step on a declared closed date is returned as exactly 0. Measured on a Monday/Wednesday/Friday market stall: with the shut days sent as null and nothing declared, every algorithm forecast about 45 units for a Tuesday the stall has never once opened — and the comparison could not see the failure, because the held-out Tuesdays were filled in the same way. Sent as 0 instead, the seasonal-naive baseline returned exactly 0.0 on the shut days but two algorithms were dragged to less than half the true level on the open ones. Declared, every algorithm keeps its open-day level and returns 0 on the shut days. What the calendar reaches A closure that repeats on a fixed weekly cycle is closed_weekdays, for a daily series only: a weekly or monthly row has no weekday, so at those cadences name the closed periods by their row date in closed_dates instead. A one-off closure — a refit week, a flood, a holiday the business did not open on — is closed_dates. Do not narrate the shut dates in your reply and leave the numbers standing; put the dates in the field, so the model is fitted around them and the forecast returns 0 on them. A holiday is not always a closure. For each one, ask the operator one question: were you open that day? Shut goes in closed_dates, and the forecast returns exactly 0 on it. Open, but trading at an unusual level — a Christmas peak, a bank-holiday lull — goes in holidays, and the model learns the effect from its past occurrences (case 3 of the driver section below). A date cannot be both: one named in closed_dates and in a holiday is refused. A standing weekly closure still governs the days it shuts inside a holiday's window: a holiday that l…

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/forecast-coach-us-west-pmcp-run"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/forecast-coach-us-west-pmcp-run)