Data Intelligence, Business Analytics
Spare parts forecasting poses a big challenge to all forecasters who are in charge of producing a good enough forecast. The reason is that the demand is irregular and a big proportion of the demand data is zero. The traditional ARIMA approaches don't work well here. Could you suggest other approaches, based on your experience or research, to this kind of demand forecast? Thanks.
If the demand is very irregular you are not going to get a great forecast. I would suggest trying exponential smoothing, or doing some bootstrap sampling. If there is a lot of months with zero demand, you could presume that the demand follows a MTBF distribution. What is the industry?
First of all, what is a MTBF distribution? Secondly, I am looking at spare parts demand for electronics devices such as smart phones. Thanks. Thirdly, when you say bootstrap sampling, how do you take into account of the autocorrelation in demand sequence? Thanks.
Sorry. I meant Mean Time Between Failures. Your "failure" would be demand. Try it using a Poisson or Weibull distribution.
The autocorrelation is a problem but you could simulate that by including the previous period in the bootstrap Of course this is not an exact method. Or you could develop a conditional distribution which would do the same thing.
So, are you suggesting to decompose the problem into two sub-problems: 1) model the distribution for MTBF using Weibull distribution, and then 2) model the demand that follow right after the MTBF using Poisson distribution? Please verify. Thanks.
No what I meant was that you divide your data into maybe x rolling 12 month periods.
So period x is autocorrelated with period x-1. Then if you are forecasting 3 months in advance, you could take 1000 bootstrap samples of 3. You could either use your actual data, as Joseph suggested, or use a hypothetical distribution such as the Poisson.
Thanks. This spare part forecasting is for high tech gadgets which have high "product innovation speed". In other words, it does not have enough data points/actual demands for its spare parts. Do you still think that Poission works in this case? Or I am just out of luck no matter what I try.
If you do not have enough data, you have to assume some sort of distribution, or try to do a simulation based upon your assumptions and model the different scenarios.
The Poisson distribution is likely the best way to approach this problem. Instead of looking at it as a timeline of when you’re going to need parts, look at it as a problem of how many parts to keep in stock. The time series is bound to get screwy, ignore it if you can.
Looking at MTBFs is conceptually sound, but I think it would present problems in operation. Let me offer an example as to why: A typical name brand SAS hard drive has a spec MTBF of 1.2 to 1.6 million hours, or 137 to 182 years. This will give a spec Annual Failure Rate (AFR) of about 0.5% of drives per year. MTBF math is kind of weird. (A lot of MTBF math also assumes a constant failure rate, when in reality it’s a curve where the failure rate starts relatively high, lowers, and then increases over time.) Those are the specs, but the actual SAS drive replacement rate per year in my company’s datacenter is around 5%. Looking at your own observed AFRs rather than specs is important if you’re going to take an approach based on them.
If you’re trying to forecast a spare parts budget, exponential smoothing on properly scaled past expenditures is probably a good place to start.
A couple of thoughts...
Techniques: Croston's method is quite popular. You can also consider using Monte Carlo simulation.
Process: I would assume that you need a forecast for some other type of planning (such as inventory/procurement planning to support service/maintenance operations). If so, you might consider adjusting your forecasting time period. For example, you might have no reliable pattern on a weekly or monthly basis for a spara part, but a very stable demand of 3 units per quarter. If that's the case, then you could make your planning decisions on a quarterly basis instead of weekly or monthly.
Value: As a follow-on to my Process point, the forecast might play a lesser role than your inventory/procurement policies. For example, if the part is inexpensive (and small to store), then you might decide to hold a larger safety stock to account for your foreast uncertainty. If the part is expensive, you have to decide if it's critical to actually hold it in stock or can you deal with a lead time and just order it on demand.
-Brian Lewis, PhD
Spare Parts forecasting is an important for supply chain management. The characteristic of demand for spare parts inventories is difficult to predict because of being trend, seasonal, cycle components and also irregularity in the data exists; traditional methods are not good in predicting those components. Unobservable Component Model (UCM) will do the best as it is capture all the above decomposed components of time series.