How it works

Metzger Planning turns your raw Excel sales history into clean forecasts and weekly purchase orders, running 100% locally on your PC.

AI agentic control

Control your entire planning workflow from a local AI agent like Claude Code. Our custom CLI exposes every step of the app, so the repetitive parts of a planning cycle run headlessly, on your machine, without anyone opening a terminal.

Claude Code

Test a scenario with $13M revenue next year and $8M during the Black Friday / Cyber Monday sale. Check the shares against history first, then run it and compare against the base plan.

Checked the DTC shares at L2 against the last three years. Outerwear is running about four points light for November against its own history, so I left the seeded mix alone and flagged it rather than fitting the target to it.

Set both targets, ran the forecast, allocated the buy and saved the comparison. FY27 Upside lands 3.1% above base on units and pulls 30,840 units of buy forward to the week of 26 April. Both plans are open in Scenarios if you want the pivot.

Other things worth asking it

Re-forecast Q3 as if it were 1 July, and tell me which models would have won.
Which styles fall under six weeks of cover on the current buy, and what would it cost to fix them?
Rebuild last week's plan with the Ecom vendor lead time raised to 12 weeks.

The app and its CLI run on your machine. The agent is your own Claude or Codex subscription, so that conversation is covered by your agreement with Anthropic or OpenAI, enterprise retention terms included. None of it reaches us.

Data input

Plug in your existing Excel sales files. No API integrations, custom database setups or IT tickets required.

Data smoothing

Automatically clean stockouts and data entry errors without wiping out genuine promotional spikes like Black Friday.

Forecast engine

Runs a hybrid ensemble of 11 statistical methods and deep-learning foundation models, including Chronos-2. Automatically backtests and scores every one of them to pick the most accurate fit.

Three accuracy leaderboards for the Ecom channel - short, medium and long horizon - each ranking seven models on score, SMAPE, MAPE, WAPE, bias in percent and in units per month, and RMSE. A different model wins each one: HW_mul leads the short horizon at 13.1% WAPE against a seasonal naive's 16.3%, SARIMA leads the medium, and Prophet leads the long at 12.9% where the seasonal naive falls to 38.6%. The ensemble blend behind each horizon is named above its table, and the three differ.

Every model is scored on weeks it never saw, and a different model wins at a different horizon. Chronos-2 leads fev-bench across 100 real-world tasks at a 90.7% win rate, ahead of every statistical baseline and every other foundation model.

Backtest validation on a log scale, so eight years of growth and every model read at once: actual demand as a dark slate reference line with every model traced across the rolling origins from 2019 to 2026 and the ensemble among them. The models track the actual closely through the ordinary weeks while the promotional spikes stand clear above them. A legend names each series; the y axis switches between linear and log, and the horizon being validated between short, medium and long.

Forecast results

Review and adjust your numbers on screen or in Excel. An override you type holds, and the model's own baseline is kept underneath it rather than overwritten.

Disaggregate

Control the assumptions that split top-level, or middle-layer, demand down through your product hierarchy: channels, categories, styles, sizes and stores.

The shares editor, two panes with history on the left and the forecast on the right, split down the middle so the same weeks of last year sit beside the weeks being planned. The top pane holds the disaggregation node - every collection under its gender, each with its own share of the channel total. The bottom pane drops to individual styles, each with its product photo, its start and end dates, and its own share of the collection above. Sale periods keep their own columns beside the months - Summer Kickoff and BFCM in the history, Spring Event and Summer Kickoff in the forecast - and a legend under the grid names every fill. A week whose size run was mostly out of stock is not marked beside the number but printed in orange, so the figure you should not trust says so itself.

Purchase orders

Convert demand into exact weekly PO recommendations based on supplier lead times, minimum order quantities and target weeks of supply.

The purchasing inventory projection: a week-by-week walk-through of opening stock, inbound, buys, transfers and demand down to unmet demand, ending units and weeks of supply. Across the total, stock opens at 144,160 units and draws down week by week while cover falls from 13.0 weeks against an 8.0 target; a 30,840-unit buy lands on 26 April rather than arriving with everything else, and the demand the plan could not cover in time is called out on its own row, climbing from 225 units to 3,076. A legend names each fill: buy, transfer, under target, ran out.

Amber where cover falls under target, red where stock runs out.

Scenarios & BI

Run side-by-side plan comparisons with native pivot tables and interactive charts, then export back to Excel for reporting and analysis.

The built-in BI dashboard: an Excel-style free-form pivot with a field list and drop zones for Filters, Rows, Columns and Values. Two saved plans, FY27 Base and FY27 Upside, sit one per row beneath each gender with their months side by side - Mens runs 3,676,662 units on Base against 3,791,621 on Upside - and the chart above shows the eight series holding together through the history and then separating in early 2026, where the two plans diverge.

Two backdated re-forecasts on the Home screen, Q4 and Q3, each scored as a dense pivot: channel and metric down the left, periods across. Under Q4 Re-forecast, Ecom forecast 121,595 units for 2025 Q4 against 134,693 actual, a -9.7% bias at 14.2% SMAPE and +52% over a seasonal naive; the shaded Total row sums the three channels to 16.5% SMAPE and +38%. Every FVA figure on both cards is positive. A part-finished period splits in two: the landed weeks are scored, and the PLAN columns carry the projection and score nothing, so forecast and actual can always be read as a pair.

Every saved plan keeps its own accuracy score, measured on weeks it never saw.

Goal seeking

Enter a revenue or gross margin target and the engine solves the disaggregation against it, back to the unit sales, margins and prices that reach the number.

Key takeaways

Zero cloud dependency

Works directly with the Excel files you already have, wherever they sit: a local drive, a network share, or a SharePoint, OneDrive or Dropbox folder synced to your PC. Forecasting runs on your machine and the app has no server of its own, so a file in a synced folder stays inside the tenant it was already in. Nothing is sent anywhere new.

Full auditability

Every saved plan keeps the assumptions it was built on and the accuracy it later scored, so you can compare what you assumed against what actually landed, run against run.

Ready to plan with confidence?

Reduce waste. Catch the opportunities. Put your assumptions on screen so the room can agree on them.

See pricing