About 10 years ago, I started work on a set of supply chain network optimization projects: finding the optimal placement of factories and warehouses to minimize cost of production, warehousing and transportation. The optimization work is analytically hard, but much of the challenge is in getting clean, complete data prior to modeling: for transportation rates in particular, how do you get reasonable rates for lanes you have no history on ?
Let’s say we’re exploring the possibility of putting a new facility in Indiana. What does it cost to get from Indianapolis to Los Angeles or to Portland? Well, we don’t know, we’ve never done it. The transportation department could “guess” a rate but just how wrong could they be? They could ask carriers for rates but as this is not (at least yet) for real demand how accurate would that be?
(As an aside, optimization models are superb at finding data errors that make costs too low. Tell the model that you can move goods for free on a lane and you’ll find that lane gets used - a lot.)
One approach to this problem just uses an estimate of freight cost based on mileage. We do at least know the mileage between 2 points. That should be a reasonable basis for optimization… right ? Well no, it’s not that easy.
Transportation rates are driven by mileage but by many other things too. As you might expect, much of this is related to supply and demand:
At this time I was working for a large manufacturer with a fairly large fraction of $1 billion in freight. They did not source product everywhere or send it everywhere but had reasonably broad coverage. To enable quick turnaround of rates for modeling I used the historical data we did have (freight rates by lane), a few simplifying assumptions and an econometric model (multivariate regression) to both quantify the impact of these factors and predict freight rates for all the lanes we did not have.
This model was very successful, the model explaining over 90% of the variation in the historical data and we used the results for some time in optimization modeling work. But (and it’s quite a big ‘but’) as history did not have as broad a coverage as we really needed I had to make some assumptions to make this work. We really needed more data, much more data to do this well.
Sometime after, the folks at CHAINalytics introduced me to their “Model Based Benchmarking Consortium”. The idea was to pull together freight data from a group of companies, pool the data in one database and build econometric models to explain what drives freight costs. I had to like the approach :-) : they collected much more freight volume than I could and in doing so built better models. The consortium has continued to grow since then, the modeling approach is continually refined and they can routinely test new ideas around what drives freight costs like:
If you're in the market for Supply Chain Network Optimization, talk to your analytic providers about how they source and validate the cost data that feeds into the model. Once you are looking at the model results it's all too easy to forget about how issues with data inputs or model structure could mean the models are lying to you. Getting freight-rates wrong can cost you a lot of money.
I have no financial interest in whether you join the CHAINalytics consortium, or anything similar, but I do really like this model based approach to benchmarking. If you can't join such a program you can still do a lot better than guessing by "sucking on your own fumes" and building similar models using your own data - I did.
How do you handle this problem? Have you found a better solution?
One approach to this problem just uses an estimate of freight cost based on mileage. We do at least know the mileage between 2 points. That should be a reasonable basis for optimization… right ? Well no, it’s not that easy.
Transportation rates are driven by mileage but by many other things too. As you might expect, much of this is related to supply and demand:
- There are areas of the country that are net importers (lots of freight going in, very little coming out). Freight-carriers charge more for freight going to these locations as they know they will have difficulty getting a paid load coming back out and offer big discounts to get a paid load coming back out.
- Similarly there are areas that are net exporters.
- Some products require specialist equipment (like refrigerated trailers) that is essentially a separate market with its own structure of supply and demand.
- There are some (though surprisingly limited) economies of scale at a lane level
At this time I was working for a large manufacturer with a fairly large fraction of $1 billion in freight. They did not source product everywhere or send it everywhere but had reasonably broad coverage. To enable quick turnaround of rates for modeling I used the historical data we did have (freight rates by lane), a few simplifying assumptions and an econometric model (multivariate regression) to both quantify the impact of these factors and predict freight rates for all the lanes we did not have.
This model was very successful, the model explaining over 90% of the variation in the historical data and we used the results for some time in optimization modeling work. But (and it’s quite a big ‘but’) as history did not have as broad a coverage as we really needed I had to make some assumptions to make this work. We really needed more data, much more data to do this well.
Sometime after, the folks at CHAINalytics introduced me to their “Model Based Benchmarking Consortium”. The idea was to pull together freight data from a group of companies, pool the data in one database and build econometric models to explain what drives freight costs. I had to like the approach :-) : they collected much more freight volume than I could and in doing so built better models. The consortium has continued to grow since then, the modeling approach is continually refined and they can routinely test new ideas around what drives freight costs like:
- How does the structure of your fuel surcharge program impact your non-fuel costs?
- Is it better to have a “core-carrier” program with fewer carriers hauling the majority of your freight or manage a wider diversity of carriers?
If you are part of the CHAINalytics consortium, you get access to their results and an opportunity to both benchmark your own freight AND to quantify and test which strategies may help you drive costs lower. If you are not part of the consortium, or something like it, perhaps you should be?
My take
I have no financial interest in whether you join the CHAINalytics consortium, or anything similar, but I do really like this model based approach to benchmarking. If you can't join such a program you can still do a lot better than guessing by "sucking on your own fumes" and building similar models using your own data - I did.
How do you handle this problem? Have you found a better solution?
This is a nice post and you hit on an important issue. Since transportation is an important part of a network design study, you need to put some thought into how you build your transportation rates. And, as you mention, you know the lane rates that you use. But, the value of network design is to consider new lanes.
ReplyDeleteUsing a source like CHAINaltyics is a good idea.
In our network design book (http://networkdesignbook.com/), we talk about other methods like using regression analysis, using "retail" rates and adjusting them, or even using some high level rates.
One thing that we stress in the book and modelers should keep in mind is the level of accuracy needed. For example, if you are building a model with demand for the next 3 years, your demand forecast is not that accurate. Therefore, you may not need to precisely define your lane costs-- good estimates may be good enough.
Mike - I'm looking forward to reading your book !
DeleteI do agree with you that absolute accuracy is not needed, my caution to more rookie modelers than yourself though is to ensure that any estimates they make are not biased when mixing estimates with actual rates.
For example: a transportation manager may offer very conservative (high) estimates on new lanes to ensure that should a lane be chosen they have a freight cost budget they can work within. These new lanes are then less likely to get used in a model. If all new lanes rate estimates are for new locations, they will all look less attractive than they should.
I have similarly seen instances where estimates for a lane rate only included linehaul rate but the actual lane costs included all accessorial charges too. It's a dumb mistake, but it introduces a significant bias between old and new. These new lanes got used a lot by the model :-)