Forecast Machine by Crest
Forecasting that works the way your business does.
39 followers
Forecasting that works the way your business does.
39 followers
Forecast Machine learns how your business works; then predicts what comes next. Give it time-series data, business constraints, promotions, seasonality, and external drivers. It can auto-fetch commodity prices and other signals, attach contextual datasets, and learn how they interact with your data. The result: business-aware forecasts across dates, products, and locations, without rebuilding models or manual pipelines.
@sankshit Hi Sankshit, congrats on the launch. How does this deal with non commodity use cases?
@zolani_matebese Forecast Machine is built to absorb signals beyond raw numbers, including business inputs that aren’t directly quantifiable and often live as tribal knowledge.
It works with any time series influenced by real-world rules, events, and context such as promotions, capacity limits, calendars, policy changes, or operator judgment. If commodities aren’t relevant to a use case, you simply don’t add them, nothing else changes.
The system still produces business-aware, explainable forecasts that respect real operational constraints instead of treating history as the only truth.
Hey @sankshit, this sounds genuinely useful. Most forecasting tools look clean on charts but break the moment promos, supply constraints, and pricing shocks show up.
Curious what your strongest real-world use cases are (inventory + replenishment, demand planning, procurement with commodity price swings, staffing, etc.) and what the business wins look like beyond accuracy, like fewer stockouts/overstock, faster planning cycles, or better decisions. Would love an example of a company outcome you’ve driven with this.
Thanks @swarnimazen , that’s exactly the failure mode we built this for.
Accuracy alone doesn’t help once promos, supply caps, or one-off events hit.
The strongest real-world use cases so far have been inventory & replenishment planning and procurement planning, especially where forecasts directly drive PR/PO decisions.
In a few cases, it’s also been used for capacity and throughput planning (vendors, warehouses) where hard constraints matter more than curve fit.
The biggest business wins haven’t been “better RMSE” — they’ve been:
forecasts go from “one-off analysis” to a repeatable workflow, cutting days of manual rework.
fewer stockouts and fewer post-promo leftovers.
Reduced emergency buying (which is usually the most expensive buying).
fewer last-minute stockouts caused by ignoring constraints
Better what-if conversations (promo on/off, capacity up/down, lead time changes) without rebuilding models.
and much faster planning cycles because teams don’t have to keep re-explaining context every time they rerun a forecast
One concrete outcome:
teams were able to move from spreadsheet-driven planning to a repeatable forecast → plan flow, with assumptions and constraints made explicit instead of living in people’s heads. That alone reduced rework and back-and-forth more than any marginal accuracy gain.
We’re very deliberately optimizing for decision quality and trust, not just prettier charts.
@sankshit This app looks interesting but I don't understand which data should I upload and for which period of time to get some predictions that make sense?
@seacat
You don’t need to upload “everything.” Start with the single metric you actually make decisions on (for example: daily or weekly demand, sales, orders, or throughput).
Even if data is on daily / transactional level, you can use the resample feature to aggregate it on weekly or monthly level.
As a rule of thumb, 6–12 months of history is enough for usable short-term forecasts, and 12–24 months helps if seasonality or promotions matter. Recent data matters more than sheer volume.
From there, you can layer in context you already use today, promos, holidays, capacity limits, stockouts, lead times, or even judgment-based rules. These don’t need to be perfectly quantified.
The goal isn’t to build a perfect dataset upfront; it’s to produce forecasts that still make sense once real-world constraints and context are applied.
Example input dataset
A simple time series with date + quantity sold, used to predict future demand.
Adding context
You can add custom instructions to explain past promotions, one-off events, or known constraints.
If you already track promos or events as time-series data, you can attach that dataset as well so the engine learns how demand behaved during those periods.
Let me know if you need some assistance in running your first forecast.
@sankshit Thank you!