Before you sign, the system decides your past weeks again, using what was known each day, and we compare it with what you did.

For companies that replenish many shops, customers or warehouses and do not want to rely on a demo with someone else's data. You give us the history of sales, stock and shipments; the system repeats each week of that period without seeing the future, and we show you side by side the days at zero stock, the excess units and the shipments it would have made against those you made. If it does not beat what you do, we tell you then.

Book a 30-minute demo How it works

Read-only connection · Tested on your own data first · Data hosted in the EU · Demos in English or Spanish

Screenshot of Arume's replenishment software: weekly shipments per location, items at zero, sales and forecast, and an urgent stock alert

The short answer

A simulation on historical data (a backtest) is a test where the inventory system decides your past weeks again using only what was known each day, and its result is compared with what your team did and with a simple rule: days at zero stock, excess units, number of shipments and demand served. It is done before you sign, on your data, with the measures fixed in advance. One or two months in shadow mode follow.

By Manuel Gómez González, founder of Arume · Reviewed on 4 November 2026

What it does

Your past weeks, decided again

We pick a period you know well, such as the last back-to-school season or the last summer. The system starts with the stock you really had at each location on day one; each week it forecasts, proposes shipments, treats them as arrived with their real lead time and moves on to the next week, subtracting the demand that really happened. At the end there are two histories of the same period: yours and the system's.

With what was known each day

The rule that makes the test valid: to decide on 4 September, the system uses only data up to the 3rd. Not next week's sales, not the stock that arrived later, not a model trained on the whole year. It is retrained at each date with what existed then, as it would happen live.

Compared on what matters to you

Days at zero stock by product and location, excess units at the end of each week, number of shipments, demand served and idle stock in euros. The measures are fixed before any result is seen, and all are shown, including those that come out worse.

With a simple baseline beside it

Besides what you did, a simple rule is simulated (restock to cover two weeks of average sales) under the same conditions. If the system does not improve on that rule, it is not worth it, and you will see it in the table.

The real lead times and constraints

The simulation uses the actual lead times of each shipment and supplier measured from your history, the days a truck leaves, minimum pack sizes and the stock that was in the central warehouse. A system that proposes sending what was not there has won nothing.

What you receive

A report with the comparison, a table by product and location to review in detail, the list of weeks when the system would have done worse than you with the reason, the data quality report and the check of the forecast. It is yours even if you do not continue.

Then, shadow mode

If the simulation convinces, one or two months with fresh data: the system proposes every week, your team decides as usual and notes what it accepts. At the end we compare again. It is the test that cannot cheat, because nobody knows the future.

What is a simulation on historical data?

It is a test in which the inventory control system decides again, week by week, a past period of your business using only what was known at each moment, and what it would have done is compared with what you did. It is also called a backtest. It answers the only question that matters before buying: with my products, my locations and my lead times, would it have done better?

It is not a demo. A demo shows how a program works with data prepared to look good. The simulation uses your data, with its gaps and odd weeks, and it can come out badly. That is why we offer it before you sign: if it comes out badly, better that you know then.

How do companies usually decide whether a replenishment system works?

Normally by watching a presentation, a forecast chart that follows past sales almost perfectly and a case study from another company. None of the three says how it will work with your catalogue. The perfect chart is often made with a model that has already seen the data it claims to predict, and the other company's case is that company's.

Take a chain of eight stationery shops that wins or loses its year between late August and the third week of September. What it needs to know is concrete: would the system have sent notebooks and backpacks on time to Alicante and Vigo, would it have left too much paper in Logroño in October, and how many shipments would it have made? Only its own history can say.

What data do we ask for?

Just enough to reconstruct what happened each week: what was sold, what there was, what was sent and how long it took to arrive. Almost everything comes from an export of whatever you use, be it Odoo, Business Central, Sage, Holded, SAP Business One, your online shop or an Excel file. What is missing is reconstructed or noted as a limitation.

We do not ask for customer data. For replenishment, product, location, date and quantity are enough. If your export mixes it in, it is removed before loading.

Sales by product, location and day, for at least a year. With two, seasonality is seen twice.
Stock today and movements in and out, or daily stock if you keep it.
Shipments and transfers made, with departure and arrival dates.
Lead times and minimum quantities for each supplier, and the days a truck leaves.
Promotions and price changes with dates, if they exist.
Codes that replaced others and openings or closures of locations.

How do we avoid looking at the future?

It is what separates a valid simulation from a useless one. If, to decide the shipment of 4 September, the system uses, even by accident, something that happened later, the result looks better than it would live and the test proves nothing. The forecasting textbook by Hyndman and Athanasopoulos describes the principle as time series cross-validation: the training set consists only of observations that occurred prior to the observation that forms the test set.

In practice the future leaks in through unobvious places, and each one is closed with a rule written before we start.

The model is retrained at each simulated date with data before it, not once with the whole year.
The starting stock is what you really had on day one, rebuilt from movements; from there it evolves with the system's shipments and the real demand, not with yours.
Lead times are those measured up to that date, not the average of the whole period.
Demand on days you were at zero is estimated before starting and is the same for all three histories, so nobody gains from your stockouts.
The catalogue is the catalogue of then: a product listed in October does not exist in September.
Parameters (service level, days of cover) are fixed before simulating and not touched after seeing results.

What is compared, and against what?

Three histories of the same period are compared: what your team did, what the system would have done and what a simple rule would have done, restocking to cover two weeks of average sales. The simple rule is there for a reason: the M5 forecasting competition showed how easily elaborate methods can disappoint when they are not validated honestly. A system that does not beat a two-line rule is not worth having.

The measures are fixed before anything is seen and all are shown, including the one that comes out worst. That way nobody can pick afterwards the one that looks best.

Days at zero stock by product and location, and demand that could not be served.
Excess units at the end of each week and idle stock in euros.
Number of shipments and lines, which is what it costs to move product.
Demand served over the total, by family and by location.
Whether the forecast delivers what it promises: when it said a 10% risk, how many times it happened.

What do you receive?

A report you can read in half an hour with the comparison of the three histories, and behind it a table by product, location and week for anyone who wants the detail. It includes the list of weeks and products where the system would have done worse than your team, with an explanation, because that is the first thing you will want to look at.

It also includes the data quality report (gaps, duplicate codes, lead times that do not add up) and the check of the forecast range. If in some period the system does worse than you, it appears in the report just as what it does better. Everything delivered is yours even if you decide not to continue.

After the simulation: shadow mode

The simulation says how it would have gone with the past. Shadow mode checks it with the present: for one or two months the system reads your data each week and proposes shipments, your team decides as usual and notes what it would have accepted. Nobody knows the future, so there is nothing to tune to a known result.

At the end the measures fixed at the start are compared again. If the result confirms the simulation, the system moves to production connected to your software, with the person who decides approving each shipment. If not, we look at where the two tests diverge and why before going on.

What does it not do, and how do we start?

The simulation does not guess what is not in the data: if a large customer joined in September and nobody noted it, the system did not know and your team did, and the report points it out. Nor does it promise that the future will resemble the past; that is why shadow mode exists. If there is any personal data in what you send us, it is handled under a data processing agreement and deleted at the end if you do not continue.

The first step is the free analysis: you send an export of sales and stock and we return what ran out, for how many days and what is sitting idle. If there is room for improvement, we fix the period and the measures with you and the simulation starts. The 30-minute video demo, in English or Spanish, lets you see it first on a sample company.

Frequently asked questions

What data do you need for the simulation?

Sales by product, location and day for at least a year, stock today and the movements in and out (or daily stock if you keep it), the shipments or transfers you made with departure and arrival dates, and supplier lead times. Promotions with dates too, if there are any. No customer data.

How do I know you are not cheating by looking at the future?

Because the rules are fixed in writing before we start and the method can be reviewed. At each simulated date the system uses only earlier data and is retrained on it. It is the evaluation described in forecasting textbooks as time series cross-validation. We give you the log of which data was used each week.

What if the system does worse than what we did?

We tell you and the report shows it, with the weeks and products where it lost and why. Sometimes your team knew something that was not in the data (an incoming customer, a building project), and that can be added. Sometimes what you do already works well, and then there is no point continuing.

How long does it take?

It depends mostly on the state of the data. With a clean export of one year, the simulation and report take a few weeks. If daily stock has to be rebuilt or changed codes matched, it takes longer, and the data quality report tells you how much before we start.

What is shadow mode?

The second test, after the simulation and before real use. For one or two months the system receives your data each week and proposes shipments; your team decides as usual and notes what it would have accepted. At the end we compare what it proposed with what happened. With fresh data nobody can fit anything to a known result.

What do you do with our data?

Use it only for the simulation, on servers in the European Union, under a data processing agreement if any personal data is involved. At the end, if you do not continue, it is deleted and we confirm in writing. The report and results table are yours.

Is it the same as the free analysis?

No. The free analysis is a first look from one export: which products ran out, for how many days and how much money is idle. The simulation goes further: the system decides each week of the period and is compared with what you did. The usual order is the analysis first and, if there is room to improve, the simulation.