Schedule Reliability Edges Up as Diversions Stabilise

The Analysis

Sea-Intelligence reports that global schedule reliability edged up in March, and it gives the reason: Cape of Good Hope routings have settled into a predictable pattern. Asia-North Europe posted the largest gain. Then comes the sentence that decides what any of us should actually do next, which is that absolute levels are still below pre-diversion norms.

My reaction to a line like that is not relief. It is: which number moved, and by how much. Edged up is a direction, not a quantity. Predictable pattern is a claim about the spread, not about the average. And still below pre-diversion norms is a bill somebody is paying every week, in safety stock. Show me the number, not the feeling.

So we do the only conversion that pays: reliability into days of buffer, buffer into cash. I will spoil the result now, because it runs opposite to the headline. On the assumptions laid out below, one percentage point of reliability is worth about 1.6% of your buffer, which on a 1,200-FEU book is 0.10 days and roughly 16,600 dollars of cash. The improvement is not where the money is. The gap to the old norm is, and so is the length of the tail.

Start with what schedule reliability actually is, because most planning systems store one number when the thing has two. Call p the share of vessels arriving inside the window your plan assumes. Call L the average lateness of the ones that miss it. p tells you how often you get hurt. L tells you how badly. A buffer sized from p alone is a guess, and so is one sized from L alone. Nearly every inventory argument I have sat in has been two people each holding a different half of that pair and not noticing.

Write one sailing's delay as a variable that is zero with probability p and roughly L with probability 1 minus p. Mean delay is (1 minus p) times L. Standard deviation is L times the square root of p times (1 minus p). Add 1.65 standard deviations for about 95% cover and you get B = (1 minus p) times L, plus 1.65 times L times the square root of p times (1 minus p). That is a two-point approximation of a distribution none of us has seen, and I will not dress it up as more. It is good enough to rank decisions, and ranking decisions is all we need it for.

Two properties fall out of it immediately, and both are worth more than the headline. B is exactly proportional to L: take 10% off average lateness and the buffer drops 10%, whatever p is doing. And B is far less sensitive to p than people assume. That is the part we are about to price.

Before the numbers, the bookkeeping. The report gives us an ordering, not levels. Reliability rose, Asia-North Europe rose most, everything is still below the pre-diversion norm. It does not publish the points. So we plan in ranges and we label every input as ours. Assume the pre-diversion norm was p = 80%. Assume today sits somewhere between 65% and 75%, a gap of 5 to 15 points. Assume average lateness L of 6 days as the base case and 12 days as the long case. None of that came from Sea-Intelligence. Replace it with your own arrival file the day you have one.

At L = 6 days the buffer runs like this: 5.2 days at 80%, 5.8 at 75%, 6.3 at 70%, 6.8 at 65%, 7.3 at 60%. At L = 12 days every figure doubles: 10.3, 11.6, 12.7, 13.6, 14.5. Read down the column and then across, because the across move is the bigger one. Going from a six-day average lateness to a twelve-day average doubles the inventory you need at exactly the same reliability. Nobody argues about that in a tender meeting, and they should.

Now price the move itself. Assume March delivered one point, from 69% to 70%. The buffer goes from 6.44 days to 6.23 days. That is 0.21 days for two points, so roughly 0.10 days per point, which is 1.6% of the buffer. Assume five points instead and you get 0.55 days. Assume the entire gap to the pre-diversion norm closed, all ten points from 70% to 80%, and you get 1.18 days, an 18.6% cut. That is the whole prize for a full recovery and it is still barely a day.

Here is the worked example, assumptions stated one by one. 1,200 FEU a year into North Europe, which is 100 a month and 3.29 a day. Landed value 46,000 dollars a box. Carrying cost 22% a year, built from 8% cost of capital, 8% warehousing, 4% obsolescence and 2% insurance. Per FEU per day that is 46,000 times 0.22 divided by 365, or 27.7 dollars. Every one of those is our number. Yours will differ, and the shape of the answer will not.

Base case p = 70% and L = 6 days gives 6.3 days of buffer. Safety stock is 6.3 times 3.29, about 20.8 FEU. At 46,000 a box that is roughly 958,000 dollars standing still, and the annual carrying bill is 22% of it, about 211,000 dollars. Say that out loud in the meeting. That is what the gap to the pre-diversion norm costs before anybody gets to talk about the improvement.

One point releases 0.10 days times 3.29, which is 0.33 FEU, about 16,600 dollars of one-time cash and 3,650 dollars a year of carrying. If somebody wants to cut safety stock off one monthly print, that is the size of the decision. Three thousand six hundred dollars a year. It will not survive the first late vessel of the quarter.

Closing all ten points to the old norm releases 1.18 days, 3.9 FEU, about 178,000 dollars one-time and 39,000 dollars a year. Eleven times the one-point move, and still only a fifth of the buffer. At L = 12 days the same ten points are worth 2.35 days, 7.7 FEU, about 356,000 dollars. The lever is slow, and it is slow in a way that scales with lateness rather than with the index. If you need a day of inventory out this quarter, it is not coming from here.

So the only question worth asking is whether the buffer is paying for itself, and you answer it at the margin. One more day of buffer costs 3.29 FEU times 46,000 times 22%, which is 33,300 dollars a year. What does it buy? Take 35 sailings a year at 34 FEU a booking with 30% late, so 10.5 late arrivals. Assume lateness spreads flat from 1 to 12 days, a crude shape that I am labelling as mine and not as data. At a 6.3-day buffer, 51.8% of those late arrivals still break cover, about 5.4 events. At 5.3 days, 60.9% break, about 6.4 events. Price one break as an expedite of 6 FEU at a 6,000 dollar premium each, 36,000 dollars a time. The missing day therefore costs about 34,200 dollars a year in extra expedites against 33,300 dollars of saved carrying.

That is not a margin. That is a coin flip with a tail on it. Which is the honest answer, and it is more useful than either slogan. It says do not cut the buffer on a headline, and do not treat the buffer as sacred either. It says the thing you are short of is a better read on L, because L is what both sides of that comparison swing on. An hour spent getting your own lateness distribution right beats a month spent reading the index.

Scale it down before you object that your book is not 1,200 FEU. The arithmetic is linear in three places, so a napkin will do. Take 200 FEU a year, which is 0.55 a day. Carrying cost is the same 27.7 dollars per FEU per day, so one buffer day costs 0.55 times 46,000 times 0.22, about 5,560 dollars a year. At p = 70% and L = 6 days the buffer is 6.3 days, so about 3.5 FEU and 159,000 dollars standing still. A one-point improvement releases 0.055 FEU, roughly 2,500 dollars one-time and 550 dollars a year. That is the honest size of the decision for a mid-size importer, and it belongs next to any proposal to re-cut parameters.

And do not stop at the buffer. What hits your cash is the whole pipeline: transit days, plus buffer days, plus whatever sits at the destination before it is booked in. If the diversion added a fortnight to Asia-North Europe transit while your buffer went up by another day and a bit, the transit half is the one that grew. Assume 38 days of transit now against 24 before the diversion: 14 extra days times 3.29 FEU times 46,000 is about 2.12 million dollars of working capital you did not used to finance. Split the pipeline into transit and buffer before anybody argues about it. The buffer is roughly 211,000 dollars a year; the transit is that, times ten; and only one of the two is a reliability question.

Which means you need your own pre-diversion baseline, and most companies have thrown theirs away. Dig out the 2023 or 2024 arrival file, or the inventory days of cover from an old board pack, and write down three figures: transit days, buffer days, and total days of cover. If you cannot find them, take the replacement assumption from your own planning system's parameter history rather than from memory, because recollection of pre-diversion service levels is uniformly flattering. The report's line that absolute levels remain below pre-diversion norms is only actionable if you know what your own norm actually was. Ours, in the arithmetic above, was 5.2 days of buffer. Yours is a number you already own and probably have not looked at in two years.

That is why the phrase in this report worth holding on to is predictable pattern, not edged up. A routing that has settled should do two things: pull the spread of arrival times inwards, and stop the occasional very long voyage. Both land in L. Because B is proportional to L, one day off a six-day average is a 16.7% cut in buffer, which is about what closing the entire ten-point gap would hand you. The stabilisation is the bigger lever, if it is real. One month does not tell you it is real.

Measure it properly or do not bother. Pull the last 90 days of arrivals on each service you use. For each sailing, take the ETA that was on the booking when you placed it, not the ETA the carrier last published. Carriers revise ETAs, and a reliability number measured against revised ETAs flatters whoever wrote it. Record the gap in days. p is the share inside your tolerance window; L is the mean gap for the rest. If a forwarder cannot give you the booking-time ETA, they cannot give you a reliability number, and whatever percentage they do quote is decoration.

The report says Asia-North Europe posted the largest gain, and that is where to start, but start lane by lane rather than with one global figure. On the assumptions above, a lane at p = 70% and L = 12 days needs 12.7 days of cover while one at p = 80% and L = 6 days needs 5.2. Same company, same week, two completely different inventory parameters. Average them and you over-stock one lane and starve the other. In most importers' own arrival files Asia-North Europe is also the trade carrying the longest transit, because it is the one that took the full detour. Check yours before you assume it.

Now the trap. These indices are monthly and they get revised, and a one-point move is worth 0.10 days, so a single good month cannot justify touching a parameter even when the improvement is real. Write the rule down and make it boring: three consecutive months of improvement, cumulative three points or more, and average lateness down 10% or more across the same window. Miss any one of the three and you hold. The rule costs nothing when you turn out to be wrong, and it stops you destocking into a quarter that turns.

There is money in the dates you send while you wait. At p = 70%, three in ten arrival dates on the customer portal are wrong the day they go out. Publish schedule plus your computed buffer, or publish two dates, the median and the p90, and say which is which. Pick one convention and hold it for a quarter. Most of the traffic around late boxes is not the lateness itself, it is the surprise, and the surprise is the cheapest part to remove.

Two booking habits earn their keep at these levels. Book the earlier of two candidate sailings on the long lanes, because the cost of missing a week is 6.3 days of buffer, not the two days the schedule implies. And split volume across two services only after you have measured both, because splitting halves your exposure to one bad string and quietly costs you the volume tier you spent the last tender winning. Assume the tier break sits at 150 FEU a quarter and check that against your own contract; below it, a reliability hedge turns into a rate increase with better branding.

Put it in the tender while you are there. Ask for the published reliability of each service over the last six months, ask for the 90th-percentile transit time rather than the advertised one, and ask for a transit-time commitment with a remedy attached rather than a target nobody owns. A carrier that will not publish a p90 is telling you it does not manage to one. That is information, and it is worth having before you sign.

Then the paperwork. Late vessels bunch, bunched arrivals are when discharge queues form and free time disappears, and a free-time clause drafted for a pre-diversion network is a clause drafted for a world that no longer exists. Take 145 dollars a day of detention and 180 a day of demurrage on a 40ft as a placeholder and check it against your own tariff. Ten and a half late arrivals at 34 FEU a booking puts roughly 357 boxes a year anywhere near that clause. Two extra days of free time is a one-line ask against a five-figure exposure, and it costs the carrier nothing on a service that is already late.

Mechanically, all of this lives in one field in your planning system and one named owner. Set transit time per service rather than per trade-lane group, and set the review cadence to monthly with a quarterly reset. Give it to one person and make the trigger explicit: the parameter changes when the rule above fires, not when somebody has a feeling. Most of the damage I have seen from reliability data has not come from bad numbers. It has come from a parameter nobody owned and everybody was allowed to edit.

Then say it to finance in their own words, because released inventory is one-time cash and carrying is a recurring line, and mixing the two is how these proposals get thrown out. On the base case: closing the gap is 178,000 dollars once plus 39,000 dollars a year. A one-point improvement is 16,600 dollars once plus 3,650 dollars a year. Put both columns side by side and let finance set the hurdle. If your hurdle for touching a planning parameter is five figures of recurring benefit, a one-point monthly move does not clear it, and that is the end of the conversation.

One more place this lands, and it is the one people forget. Sales is quoting lead times off a number somebody set two years ago. At p = 70%, three in ten of those quotes are optimistic by roughly six days on this base case. Give them the buffer number and make them choose, consciously, between quoting the median and quoting the p90. A quote that is six days optimistic and wins the order costs you more than a quote that is honest and occasionally loses one.

One distinction saves a lot of arguing. The index is a network number; your exposure is a service number. A global figure that rose because Asia-North Europe rose most tells you nothing about the two loops you actually sail on, and nothing about the week you ship. Ask for reliability and lateness broken down to service-loop level, and if what comes back is a trade-lane average, file it as a marketing number. Marketing numbers are fine on a slide. They are not fine as a parameter.

What would change my mind is specific, so name it in advance. If two consecutive months showed average lateness down 15% or more on the services you sail, with the p90 falling faster than the mean, I would cut the buffer without waiting for a third month, because that is a distribution change rather than a run of luck. If a carrier published a p90 that held for a quarter and then committed to it in the contract, I would cut further. Neither of those is what a monthly index gives you. A monthly index gives you a direction, and a direction is not a trigger.

And keep one figure in your pocket for the negotiation. On the base case a buffer day is worth 33,300 dollars a year, so a carrier asking you to accept a service with one more day of average lateness is asking you for 33,300 dollars a year, per lane. Say it out loud. Carriers quote in dollars per box; answer in dollars per day.

One thing to carry into Monday. If reliability is up and your buffer has not moved, the response is not to cut inventory, it is to go and find out which of the two numbers moved. Reliability improved. Congratulations. It improved by a tenth of a day of buffer, which is about how long it takes to work out who is going to phone the customer. What has not moved is the gap to the pre-diversion norm, and that gap is 178,000 dollars of inventory on this book at six days of average lateness and 356,000 at twelve. Go and measure L. The index will still be there next month.

↑ Back to the key points

— By Vivian Zhao