Advertisement
Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- - Joannes completed 2 years of Bioinformatics at L'Ecole des Mines before quitting in 2008 to start a company in Supply Chain
- - Supply chain is vast and varied but my definition is "mastering optionality in the presence of uncertainty and in the context of the flow of physical goods" - 2:30
- - Optionality - 2:50
- - making decisions, to order, to raise or lower price, consuming my materials to manufacture more goods
- - It's also to create or control what decisions I can make in the future - for example if I don't have a secondary supplier I am stuck when a failure occurs.
- - Uncertainty - 4:20
- - Inside a factory you can have a relatively good grasp on all of the possible failures - it's a controllable environment.
- - In supply chain we're concerned about everything between factories and warehouses -
- - Where a boat can get stuck between 2 continents and block shipping between the largest markets in the world
- - Where an epidemic can cause the largest ports in the world to shut down
- - Where a president can decide to create new barriers of trade with China
- - For example in France we had regulations put in place last year to limit e-commerce vendor's usage of discounts on elevated MSRPs
- - So if as a supply chain you were counting on the ability to take certain actions things will be significantly more difficult.
- - Uncertainty also comes in a variety of forms. Not just will there be more or less demand, but also
- - Delays are uncertain, your costs are uncertain, competitors prices are uncertain - we can not hope to one day have control over all of these variables
- - Physical goods - 6:35
- - We are playing within the constraints of the financial markets. Physical goods add many constraints that must be followed - you can't make an infinite quantity of something.
- - Supply chain converts Euros into goods, and then eventually back into Euros - making some profit.
- - Predictive Optimization is to provide automated decisions on top of these Supply Chain problems
- - In general Supply chains have been digitalized for over 40 years - we have customers where the people who designed and implemented the databases have most likely died. - 7:40
- - Systems programmed in COBOL in the 1970s
- - bar codes were popularized in the 1980s
- - The problem is not to digitalize supply chains.
- - Leaving aside the question of if they are well digitized, if we got all the benefits we wanted, with modern technologies.
- - Systems of record - 8:45
- - Transactional system are the foundation of supply chain - orders, prices, customer demand, etc.
- - These mostly originate from the 1980s with software products known as ERPs
- - This should have really been called ERM (Enterprise resource management) as there is little to no planning being done in these systems.
- - The predictive optimization piece is much more recent but some pieces have existed for years.
- - Now I will put on my software vendor hat and start disagreeing with how things are done today - 10:00
- - Its surprising - but in textbooks on the descendents of Operational research nothing works
- - Before we went into our customers they had had 5 or 6 other solutions before us that didn't work
- - For example SAP has 4 different systems for predictive optimization
- - An optimization for how to put components in an electronic device is a static optimization with known bounds in contrast the question of "should I launch this new production" - I won't have an answer until 6 months from now - so my optimization is in the context of an uncertain future
- - Stochastic Optimization
- - What are these typical operation research topics that I'm saying don't work? - 11:55
- - First Time series in demand forecasting
- - Doesn't work well for unusual demands
- - For example, where you have consumers who buy switches one at a time and construction companies who buy 500 at a time.
- - The order for 500 is announced well in advance, while the one offs are immediate, yet a time series will not differentiate between these cases and squash the demand together
- - For substitutions in fashion
- - if the wrong size is there - can't sell it, but if its a slightly different color, the customer wil likely still take it
- - Diapers and market basket effects
- - Diapers are expensive and have high brand loyalty on average
- - However it's not the loss of the diaper sale that is the most impactful - it's the fact that parents will then not buy all of their other groceries at the hypermarket if the diapers aren't there.
- - Time series also hides the fact that the future is blurry and in which direction
- - Deterministic forecasting works pretty well with consistent values
- - But a deterministic forecast of something like a soup in a supermarket with promos and demand swings of >100% it makes no sense to generate a point forecast
- - Second Safety stock - 15:45
- - Rather than thinking about products in isolation we should be answering the question
- - "I have one additional Euro of inventory I can hold - which product should I give it to?"
- - Safety stock using a Gaussian normal distribution is mathematically intelligent but not practically sound
- - At the beginning of Lokad I didn't have this view at all - for many years - 17:00
- - I read the all the books and papers from MIT, Caltech, Yale, coded all the algorithms, and nothing worked.
- - But all of my clients were happy to pay me, knowing it didn't work
- - Those clients were even happy to be a reference to other customers and tell others about how well we were doing
- - "Maybe with all the time and effort wasted if we recommend you to our competitors you will make them lose money and time" :)
- - Even though what we were using was validated by MIT, the code is right, and the customer reviewed it - it still doesn't work
- - This continues through 2012 - I'm making about $2M Euros a year, and I know none of what i'm doing works.
- - My competitors make up to Billions of $ - and they're doing the same thing.
- - With just 30 more years at 30% growth I could do the same - and yet nothing works.
- - I went through an incubator - 80% of my colleagues there have lost all their money in businesses where the customers complain while it works - for me my customers are happy and yet it doesn't work
- - A major european distributor put out a call for benchmarking forecasting tools - 19:30
- - Asking 6 different companies, some American, some European to forecast 5000 SKUs at 10 supermarkets 5 days ahead of time
- - To have the lowest error - Lokad submits just 0s on everything
- - Lokad's answer ends up being 20% better than the next runner up
- - If you forecast 0, you will replenish 0, if you replenish 0 you will sell 0, thus being statistically 100% correct with an entirely idiotic answer
- - This led to a crisis of faith - I'm following the books precisely and yet giving obviously idiotic answers
- - This led me to ask the question - what is a supply chain scientist
- - Now by rebasing my assumptions and throwing out what the literature said Lokad started doing a lot better. - 22:20
- - The first question was what are we trying to optimize for?
- - We were trying to improve percentage error reduction - but I realized we shouldn't care about error percentages, rather we want to optimize for profit
- - We're looking for Euros of error
- - Now we look to make probabilistic forecasts biased by Euros, biased in the direction of profit
- - So for example in a supermarket we'll stock more yoghurt, they have high margins, aren't bought very often, but we don't mind trashing some every once in a while since the margins are so high.
- - Now when the client shows up every week or two they will find their yogurt - so we're not biasing for the median profit/sale but the top 1 or 5% we care about to make profit
- - In operational research we have this belief that we know what we're optimizing for - however we don't really know what is optimal - 25:08
- - Supply chains are very complex, with complex software systems, processes that are not known by management
- - It is extremely difficult to count the euros and actually know what financial effect a decision will have
- - For example a newspaper may run a temporary discount. You would think there would be some curve to describe the demand-supply match against price and an increase in demand to occur. However the customers now see that you're willing to discount your product and will now wait until a similar discount is available to them before purchasing. I am cheapening my brand by providing discounts.
- - So what is the real cost of a discount? How many of my customers will now be habituated to a discounted price? That must be included in my calculation to see if the discount/promo is worth running
- - 1st order effects can be measured such as impact of promotion, but those indirect 2nd order effects e.g. consumer behaviour changes are hugely critical to take into account - 27:15
- - For example a spare parts organization of an aircraft company had a recommendation to purchase a part, the employees said absolutely not!
- - It seemed a reasonable recommendation to make - however it was a spare part for a 747 - and since the part had a 30 year life, and the 747s are being deprecated within the next 10 years there was no need for it
- - Another interesting effect in airlines is the one-way standard.
- - A plane is allowed to have a part matching the old standard, however as soon as a part passing the new standard is placed - you must only use new parts going forward.
- - This means if you have inventory of old and new parts - each time you replace with a new part- you are modifying the composition of your demand for spare parts across your flotilla
- - Just because we don't know what to optimize for, doesn't mean we can't find out - 29:17
- - It just can't be done in a top down cartesian way in which we split the problem down into constituent pieces and come up with an answer that way
- - We arrived at the need for supply chain scientists to operate under the idea of Experimental Optimization
- - To optimize you must create logic to generate a decision - you will then have people who object
- - Their reasons for objecting are typically correct
- - Use anecdotes to find the reasons it won't work
- - Then we will feedback into the dollarized/financialized decision making system to add the constraints and requirements that will meet the edge cases which are critical to that customer
- - For example with the spare parts problem we had taken into account the lifespan of the part, but not the lifespan of the plane itsself
- - In our supply chain books we tell you the demand is a Gaussian, lead times are a Gaussian - is there any way to falsify this? no, it's in the abstract, divorced from reality - 31:50
- - Our goal is to make a mathematical model , maximized for modeling reality, not necessarily mathematical simplicity
- - An important part of the process is making sure to present the results to the customer in a specific way - not just "here are your optimal stock levels" but - "where should I put my first Euro of inventory investment?"
- - 80% customers know what they're doing - so this list of prioritized investments in purchasing, manufacturing, or inventory levels is reasonable - though sometimes they may be missing something obvious
- - Working with a german MRO company
- - Retrofit and repairs are 2 different things
- - Repairs are - engineer says this part needs to be fixed, replace
- - Retrofits are - the manufacturer has some doubts about a part and requires a push to replace within a month
- - You can't mix push and pull demands together - they act completely differently
- - You push a big spike of parts - and now all of your spares for that part are synchronized
- - The supply chain scientists mission is to discover the financial optimization 36:25
- - Our customers ask us for the benefits up front - but ur answer is "we don't know yet"
- - Large industrial client - "why are you ordering like this? We are copying exactly your orders - why is there so much overstock?"
- - Turns out the receivers would change the order qty to match what was shipped in case of a discrepancy - this led to suppliers figuring out at the end of the quarter they could ship whatever and it would be successfully received and billed.
- - Since the order qty was changed there was no way to go back and state that the quantity was different than what was ordered
- - The supply chain scientist has to unknot problems like these - and for the customer's request of value up front - there's no way for us to know these kinds of problems existed before
- - Due to the diffuse and unmeasured forces in supply chain it's very difficult to have an __a priori__ view on the value. When we go through this effort of rationalization however, there are significant gains
- - But it requires a leap of faith
- - Data analysis 41:39
- - The supply chain scientist has to ask Why, Why, Why?
- - Books on Data science have a bit of a fantasy view of the job - in reality
- - ERP's have upwards of 10000 tables, each table has 1-200 columns
- - Not well documented
- - Built by people who may or may not be alive or working at the company
- - The documentation of the system does not necessarily inform the way people use the system
- - Not only do these customers have these complex ERPs, but they also have WMS, TMS, CRM, EDI, ......
- - Plus acquisitions with multiple ERPs and integration middleware
- - The Kaggle Data scientist has 3 tables with each column well documented - in reality you are working with a thousand tables and your IT team has 4 years of backlog
- - 5 years into a 3 year ERP deployment
- - The analysis is not to do advanced AI - its just to find the pertinent information
- - At Lokad we've developed a tool to uncover informational entropy - the Shannon complexity by column
- - With most of our customers we receive about one sentence of documentation per table - in the end we end up with one page of documentation per field.
- - Describing all the usage and supply chain implications
- - Finding Supply Chain Scientists 48:15
- - We need people who are able to descend through 100s of fields without losing their minds
- - Need to understand and remember the meaning and connection between the fields
- - Have to do mathematical analysis
- - Parse fantastical column names put in by geeks
- - and most importantly, when they've come up with a theoretical explanation for what the data means - figure out how to falsify their theory
- - Something like order_date can mean 20+ different things! how do you validate your interpretation
- - On Kaggle you are served up the data on a clean platter
- - Numerical Modelization 51:15
- - The median Kaggle competition has extremely opaque and complex models that are unusable in production
- - They'll use Stacking - whenever there are residuals they will stack LP with Deep learning, with gradient boosted trees to get an additional 0.1% accuracy
- - In the Kaggle Walmart forecasting competition (M5) - Lokad competed and came in 5th
- - The top 50 entrants used Gradient Boosted Trees or Deep Learning with 10s of thousands of parameters
- - Lokad won with a model with 5 parameters
- - Many of the entrants had model running times of hundreds of hours for a single locations sales.
- - Untenable for a real enterprise to spend that much effort on forecasting the next X days sales
- - One of the bigger questions our scientists have to consider is how maintainable is this model in production
- - How will we do when a customer calls us and says you have 48 hours to fix this?
- - For example when the Suez canal was stuck we had to update lead times and recalculate all of the impacts to get an answer tomorrow for many customers
- - Same with the German floods
- - The vision then for supply chain is not can you naively squeeze an additional 1 percent of error out - but how well can you adjust in 24 hours for a completely different business landscape due to a supply chain disruption/catastrophe
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement