AI: intermittent demand for spare parts
The service company spare part may not be issued for several months, and then need for several repairs at once. A permanent minimum balance answers the question “when does the rule work?” The forecast of rare demand responds to another: what need to expect on the chosen horizon and how uncertain this response is. Buying an AI server for such a task is wise after checking history, comparing methods, and determining the decision the provider will make.
Request a quote
SpectraTech can be requested AI Server Manager 256 for a separate local forecast project. The name “Superintellect” is not a proof of forecast quality, and the 256 does not specify RAM or GPU memory. The platform, configuration, software and work are selected according to the agreed volume. The current price and availability are confirmed in the quotation.
First, determine what to predict and for how long.
Suppose the service serves multiple generations of equipment and wants to assess the need for repair kits before the next arrival. To begin with, select a limited group of articles, a unit of quantity and a calendar step: for example, issuing by week. The horizon is associated with a real supply cycle, not a beautiful schedule for the year ahead. It can be different in different groups.
History should show what happened to a particular position. Zero with an open warehouse and available goods is different from missed unloading. The absence of issuance, when the details were not, does not mean the absence of a need. Returns, reclass, canceled repairs and transition to a replacement article are noted separately. Otherwise, the forecast will confidently reproduce registration errors instead of regularities of the service.
A useful source set contains the date, article, quantity, type of event, sign of availability, and known assortment changes. Planned maintenance can be taken into account if the information was available at the time of forecasting. The data generated after the repair cannot be used to predict the same period. A separate stage of preparation is needed to restore the meaning of the story and agree on exceptions with the owner of the account.
Compare methods on the same past
For rare demand, start with a simple, comparable benchmark. The official StatsForecast example examines special methods including Croston and ADIDA. The methods presented here give a point forecast; intervals require a different probabilistic approach. Therefore, a beautiful uncertainty band cannot be automatically assigned to any algorithm chosen. An example of working with rare demand.
Chronos-2 can be considered as a separate candidate: the official Python example shows a series prediction with given quantiles. It is included in comparison with the basic method on the same articles and horizons, rather than being declared the winner in advance. The model, its version, and the computing environment refer to the consistent software part of the project. Official Chronos repository.
The practical test takes place in time. Select a few past dates: for each open only the then known history, build a forecast for the agreed period, then compare with the actual issue. This principle of time windows is described in the documentation of StatsForecast; the official evaluation code of Chronos also provides for the division of a series and several evaluation windows. The test includes calm periods and periods with discharges. Time-checking, Chronos rating code.
| What the clerk sees | What needs to be explained | How to use the result |
|---|---|---|
| Long sequence of zeros | Products are available or sales are limited by lack of | First, confirm the meaning of the observations |
| Rare major rendition | Regular repair, one-time project or correction of accounting | Agree on the inclusion of an event in history |
| Point forecast for delivery time | Method error on past time windows | Comparison with a simple guide |
| Range of future values | Calculation method and actual coverage on inspection | Consider uncertainty when discussing stock |
Assess the mistakes that change the decision
For a position with many zeros, one average percent error is not enough: dividing by zero actual expense generally requires a different approach. It is useful to look at the error of quantity over the entire horizon, underestimation in periods with issuances and excessive forecast in calm periods. The results show in agreed groups so that a large number of quiet positions do not hide the failure on really important parts.
If a probabilistic method is selected, the range is also checked: how often the actual result fell into it and how wide the boundaries were. The range for individual weeks cannot be mechanically added and declared the correct interval of the total need: the desired result is checked separately. Excessively wide corridors do little to help purchase, and a narrow one can look convincing and often miss. The new model is adopted according to agreed tests, without promises of profitability or complete elimination of the deficit.
Prediction does not become an order automatically.
Forecast demand, confirmed balance, reserve for work and expected receipt are different data. In the working report, they are shown separately, with the date of the cut and sources. The supplier applies the adopted reserve policy, takes into account the minimum batch, delivery time and the possibility of replacement. After that, a purchase decision is made. The forecast itself does not rewrite the balances and does not send the order to the supplier.
To select a server, you need the number of rows, the length of the history, the frequency of recalculation, the number of verification windows and the period of readiness of the report. A small task can be done with ordinary computing resources; the GPU is considered by the selected software and control run. The AI Server Manager 256 configuration is justified by this volume, storage requirements, and work window, not the word “AI”. The quotation includes platform, history preparation, software, testing and integration within agreed boundaries.
Send an impersonal example of unloading, a list of groups, the purchase horizon and the current replenishment rule. This will allow you to offer a comparable pilot and specification instead of buying power blindly. Delivery and timing are agreed for your city in Russia or Kazakhstan, for example, Novosibirsk or Astana; local warehouse in each city is not meant.
Checking the past
- History known at the selected date
- Basic method and model on the same horizon
- Comparison with subsequent actual issuances
Working decision
- New forecast with cut-off date and range, if checked
- Separately: balance, reserve, receipts and delivery dates
- Supplies apply the policy and confirm the purchase
The forecast does not change the account balance and does not automatically send the order.
Meaningful design of the project. Future actual outputs are not available to the model at the time of the trial forecast; the interval does not guarantee the absence of a deficit.
Questions before buying
Why is the minimum balance rule not enough?
The minimum sets the adopted policy of action with a margin. The forecast assesses the future need on the chosen horizon. The utility of the additional predictive module is checked in past periods; a simple rule may remain sufficient for some positions.
Does the forecast range guarantee the right details?
Nope. Its method and actual coverage are tested on time slices. Purchase additionally depends on the balance, reserves, terms, minimum batches and the decision of the company. Interval is not a guarantee that there is no shortage.
Do you need a GPU and Chronos-2 for this task?
Nope. In comparison, simple methods and a probabilistic model can be involved. Computing resources are selected by the volume of rows, the working window and the tested software. Chronos-2 is a pilot candidate, not a mandatory bundle of software or a pre-best algorithm.
The main product in our catalog
AI Server Manager 256
Price on request. Configuration, availability, software and work are agreed in the quotation. The designation of the model does not determine the memory capacity or composition of the processors and accelerators.
Request a quoteDocumentation: Nixtla StatsForecast: Intermittent Demand, official tutorial; Nixtla StatsForecast: Cross Validation Official Teaching; Amazon Science: Chronos forecasting, official README main; Amazon Science: Chronos, official evaluation script main. These provisions relate to these editions; the applicability to the proposed configuration is specified separately.