By Simon Schalit, January 2015
When a company commissions a full power plant, industrial heavy machinery, or fleets of aircrafts or cars, it expects this investment to generate revenues for the coming years, if not decades. To ensure the return on investment, the maintenance and servicing over the long-run of said equipment is crucial and usually represents a significant part, if not the majority, of the costs of the project. In order to cover this risk, resorting to long-term maintenance/service agreements offered by the vendor party (OEM, MRO or other) has become the norm.
These agreements can take several forms. However the bottom-line is almost always the same: the financial risk of maintenance is fully or partially transferred to the vendor over a set time horizon (years or decades) for a price set at the beginning of the contract. The question then becomes: of the two parties, who can better assess this risk and get the upper hand in the negotiation? And for the vendor, how to optimize its process to maximize the margin during the contract?
Assess the risk pre-sale and live with it
Considering the financial importance of long-term maintenance contracts, and the fact that it is not uncommon for a vendor to sell the equipment itself at a huge discount while counting on the maintenance contract to generate margins, the pricing and conditions of the service are usually at the center of the negotiations between the parties.
Companies use a variety of tools and processes to estimate the costs tied to different maintenance actions that can be expected (cost of the parts to be replaced, dedicated man-power for each type of intervention, costs of service interruption…). However, though this estimate can be complex, it addresses but a small fraction of the problem. The real challenge remains: how likely are those cost-generating-events to happen at any given time and how often over the long run? If the vendor under-estimates the risk, he may end up losing money over the course of the contract. On the other hand, if the vendor over-estimates the risk, thus over-pricing its offer of service, he may end up losing the contract entirely.The reality of long-term contracts is that the final cost is very uncertain, and thus could reasonably vary over a wide range. Any attempt to come up with an “accurate” value that would try to be “right” or at least “close to the truth” suggests a real misconception of the forecasting process. There is simply not “one accurate value”; any forecasted estimate will carry a level of risk, and it is the assessment of this (financial) risk, expressed in dollars, that should be at the center of the forecasting process.
Once the maintenance agreement is signed, the vendor is going to have to live with it. However, this does not mean that the forecasting effort stops here. On the contrary, regular updates on the risk are necessary to ensure the viability of the contract. This includes:
- Short-term forecasts to optimize the resources (spare parts inventory and man-power) to be maintained to ensure appropriate response time and service level. These forecasts are short-term in the sense that they focus on a “process horizon” (or Lead Time), to ensure this process is as lean as possible.
- Long-term forecasts to refine the assessment on the risk still carried by the company for the remainder of the contract, and the calculation, should the case arise, of loss provisions. The danger with long-term maintenance contracts is that most of the costs are often accrued towards the end, while revenues are usually recognized regularly over the duration of the contract.
Limits of classic approaches to maintenance forecasting
Assessing the risk and the costs attached is a difficult task, and unfortunately, this problem is typically one where the classic approaches used by most companies work poorly.
The simplest methods, relying on the specifications provided by the manufacturer (MTBUR type of data for example) give but a poor representation of the reality as the reliability of parts is often greatly impacted by external factors (usage, environment…). In our experience, real reliability patterns have little to do with theoretical figures, especially over the long run.
More advanced classic methods, relying on traditional statistical “classic” forecasts also fail to capture the reality of the patterns encountered with spare parts. These methods rely on the assumption that maintenance forecasting is just like any other “demand” forecast and thus can be addressed using the same approach. This is unfortunately untrue. Several specificities make forecasting for maintenance hard:
- Rare events: mechanical failures are by definition rare events, so when looking at specific parts, relying heavily on models offering “smooth” patterns (much like retail top-sellers) is somewhat naive.
- Replacements in waves: the reality of maintenance is often that the disruption of service is more costly than the broken parts themselves. This is a strong incentive to replace parts in waves, rather than one at a time to avoid unnecessary downtimes. This invalidates the assumption of the different parts having “independent” maintenance patterns and, with it, most of the popular forecasting models that happen to rely on this assumption.
- Extremely high service levels expected: considering the cost of an interruption of service, the expected service levels for maintenance contracts are often extremely high, way above the range generally targeted in other industries. As an example, the cost of an aircraft on ground (AOG incident) can range up to several hundred thousand dollars per day.
- Closed-loop repair cycle: Many parts are just too costly to be thrown away. Some are sent for check and repair and then reimported back in inventory for future use. This takes the company out of the traditional “sell and reorder” scenario. Once the company has purchased the part, it can stay in inventory for a long time. This makes the decision of purchasing to increase the inventory all the more serious, as it commits the company for a long period.
However the biggest hurdle is the concept of classic forecast itself. By definition, the forecast in the classic sense is not a prediction nor a guess, however accurate it may be. It is a statistical estimate of the expected median of the demand/cost. So in this case, classic forecasts applied to estimate the overall cost of a maintenance contract would provide a value that would have, by definition, a 50% chance to be above or below the real cost. Of course, from a financial perspective, these odds are unacceptable in this situation, which renders the concept of classic forecast irrelevant. In the end, the key to generating adequate forecasts is to adopt a financial perspective on the forecasting process from the start. The objective is to rely on “forecasted scenarios” taking directly into account in the forecast the target financial cover (financial risk, service level) to be achieved, and thus the underlying financial driving forces. And that is quantile forecast.
Lokad’s Gotcha: trying to transform a traditional classic demand forecast into a financial cover by adding a safety buffer on top (usually referred to as a “safety stock” when talking about inventory) is nothing but a very inaccurate way to generate a quantile forecast.
A financial perspective on forecasting: the quantiles
Forecasting for maintenance is first and foremost a financial optimization, both on the financial risk over the whole contract and on how lean the maintenance process can afford to be while maintaining the desired coverage/service level. The higher the estimate of the costs/stock level necessary, the lower the probability of seeing this estimate over-taken by reality, but it is important to keep in mind that no estimate can guarantee a 100% cover.
These scenarios can be generated through quantile forecasts, which are in fact an extension of the classic forecasts: instead of looking for the value that has a 50% of covering the future demand/costs, quantile forecasts allow us to determine any threshold, be it 10%, 60%, 80% or 98%, within the cost/risk distribution.
Estimate of the overall costs and remaining risk
The objective is to generate forecasts corresponding to the different levels of risk the company would be willing to accept. This analysis should take the form of several simulated scenarios, ranging from the lowest acceptable cover, providing a non-negotiable floor pricing, to higher levels of cover, providing more favorable scenarios at a higher pricing.
In reality, the pricing of maintenance contracts will be driven in no small part by the client’s “willingness to pay” and the degree of competition. Therefore, the vendor is usually bound to moderate its pricing, but generating the above mentioned scenarios will allow the vendor to actually quantify the risk he is facing for a certain pricing level.
These scenarios are also particularly useful when updated during the contract to evaluate the risk on the remainder of the contract, and thus determine if provisions need to be created or adjusted, and by how much. This approach offers the great advantage to provide a quantification of the risk, thus allowing a direct financial estimate and a complete control on the prudence level to be adopted.
Optimization of the maintenance process in contract
Concerning resource/inventory optimization, the ideal situation would be to set a target service level to be achieved and calculate the minimum corresponding resource/inventory level necessary to ensure this service level. This in itself is difficult considering the specifics of maintenance contract mentioned above, but can be achieved through quantile forecasts which allow, in the same way as with the scenarios above, to directly target the desired service level and evaluate the corresponding need.
However, the reality of maintenance is often more complicated, as companies usually need to operate under a limited budget and need to arbitrage between the different parts to ensure that they get the best ROI in terms of service level per dollar invested. This optimization is made possible by generating a quantile grid, which is the representation of the results for all part types of all possible scenarios, within the range of acceptable service levels (how many parts of each type would be needed to ensure the whole range of possible service levels). This allows the company to navigate in that grid to determine the most efficient inventory to be maintained under a budget constraint.
Lokad’s Gotcha: Several systems claim to rely on “Monte-Carlo” methods. Companies should bear in mind that “Monte-Carlo” is not a magic word in statistics and should not be used as an excuse for lack of understanding of the driving forces underlying the models and lack of proper data.