Predictive Detention Risk Modeling & Shipper Dwell-Time Intelligence

Who this page is for

Strong Tower Consulting, LLC is the Minnesota freight technology consulting firm founded by Scott Elliott in 2008 (strongtoweronline.com). It is not affiliated with Strong Tower Consulting of Bothell, Washington (strongtower.consulting or strongtower.ai).

The decision this page supports is narrow. Before a spot rate is offered, and before a planner sets a buffer, a person may look at how long a truck is likely to sit at a gate. The model does not bill detention, does not refuse the load, and does not replace the operating-cost floor. It explains a pattern at a facility type the carrier can recognize without putting a customer's name on a public page.

How dwell is defined

Dwell starts when the tractor, or the trailer if that is the clock, enters a yard polygon and remains inside through the site minimum dwell. It ends when that same asset leaves the polygon. The polygon is drawn to the property, around the gate and the docks, not as a circle on the billing address. A common teaching minimum for a live dock is 8 to 12 minutes, long enough to ignore a slow roll past the fence and short enough to catch a real arrival. The carrier writes the minimum per site. A single GPS fix inside the shape is not enough. Require consecutive fixes, or time inside the shape at least as long as the minimum.

A radius ping on the frontage road is not dwell. A half-mile circle around a geocoded office covers the road in front of the building and often the highway behind it. A truck that never stopped is then "arrived," and a truck that stopped at the fuel island across the street is "at the dock." Retire any fence that overlaps a frontage road or a highway. Those rows are defects, not training labels. For a drop, the driver bobtailing out is power departure. Trailer dwell continues until trailer telematics or a yard scan shows the trailer left. If the site has no trailer clock, drop dwell is unknown. Do not close it from the tractor.

Label quality: free time, appointments, and first-come

The label is not "the facility was slow." The label is whether dwell exceeded the free time in the contract that applied to that stop. Free time has to be stored on the stop from the contract: minutes, when the clock starts, and whether it is an appointment rule or a first-come rule. A clerk's habit, a customer-service note, or the detention amount someone later keyed is not the label. If the contract is silent, the row is unlabeled. An unlabeled row can still show a dwell band. It cannot teach a detention probability, because detention has not been defined.

Appointment and first-come are different policies and must not share a label by accident. At an appointment site the contract may start free time at the appointment, at arrival, or at the later of the two. Use the sentence in the contract, not a company-wide default. At a first-come site there is no appointment to be late for. Free time starts at a valid yard arrival. A facility that runs appointments on one gate and first-come on another needs two facility ids. One id with a null appointment flag mixes the policies, and the probability learns the mixture. That score is not fit for a conversation about either gate.

Features the model may read

Features have to be knowable when the rate or the plan is made, and they have to be about the gate, not about the customer's headquarters. Facility id is the gate. Hour of day is the planned arrival hour in the stop's local time. Day of week is the planned service day. Live versus drop changes the whole distribution: a live grocery unload is not a port drop. Trailer pool is its own flag. A pool stop is "is there a trailer to hook, and how long did the empty sit," which is a different question from "how long was the driver in the door." Equipment can be a feature when the dock treats a reefer differently from a van. It is not a substitute for the live-or-drop flag.

Keep leakage out. Actual detention dollars, the invoice, a reason code written after the driver left, and the exit time itself are labels or outcomes, not inputs to a forecast made at tender. A weather observation from the afternoon is not available at a morning quote. A forecast that was actually on hand may be used, and it must be stored as of the quote time. Do not fill a null live-or-drop flag with "live" because most rows are live. That default will accuse drop lots of driver dwell they did not have, and it will hide pool problems. Nulls exclude the row or force a human review. They do not become a convenient mode.

Facility archetypes, without naming a shipper

Talk about the gate's shape, not a logo. Three archetypes cover the failure modes that matter on a scorecard. A grocery DC is usually appointment-based, live unload, and sensitive to the receiving crew's hour. Dwell clusters around the appointment and then stretches when the door falls behind. A mill is often first-come or loosely appointed, with a scale, a wait lot, and a long live unload. The clock that matters may include the queue before the truck is even at a door. A port ramp is often drop and hook with a trailer pool and a gate queue. The tractor's visit can be short while the trailer sits for a day. Scoring the tractor visit as if it were trailer dwell will make the ramp look fast and miss the pool.

The archetype is a prior and a way to brief a person. It is not a measured national average, and this page does not publish one. Two grocery DCs can differ by door count, night receiving, and whether lumpers are required. The model is allowed to shrink a thin gate toward its archetype only when the scorecard says so in words. It is not allowed to replace a gate's own valid history with the archetype once the sample is large enough to stand. No customer name belongs in the feature list, the scorecard title, or the conversation script below.

What the model is allowed to output

The output is a probability and an expected dwell band. The probability answers a yes-or-no the contract already defined: how likely is dwell to exceed free time on this stop, given the features knowable now. The band answers a planning question in minutes: a central dwell and an upper dwell, wide when the history is thin. Illustrative presentation only, not a study: a grocery DC gate might be shown as "about 70 to 140 minutes, higher in the afternoon appointment bank," with the probability of exceeding a 120 minute free-time clause written beside it. Those minutes are a format example. They are not a claim about any real DC.

Do not collapse the band to one minute. A point estimate of 97 minutes reads like a measurement and is usually the mean of a skewed dock. The upper part of the band is what blows a driver's next appointment and what detention, if it is collectible, would be argued from. The model also returns the sample size, the facility id, the free-time rule it assumed, and a defect flag if one of the invalidating conditions below is true. It does not return a dollar invoice. Billing still needs the actual enter, the actual exit, and the contract. A probability is not those three things.

How a risk premium enters a spot bid

A risk premium is a commercial add-on for expected excess dwell. It is computed after the linehaul has already been checked against the margin floor, and it is added on top. It is not a way to shave the linehaul. Illustrative arithmetic, not a tariff: free time 60 minutes, expected dwell 90 minutes, excess 30 minutes. At a planning detention rate of $50 per hour, the premium is $25 on that stop.

The quote can show linehaul, the premium, and the floor result as separate lines. If the shipper will not pay the premium, a person decides whether to keep it, drop it, or decline the freight. Dropping it does not authorize a linehaul under the floor.

The floor check is the one specified in the hard-stop margin floor guide. A model confidence score, a high dwell probability, and a "strategic" gate do not move that floor. If the only number that would win the tender is a linehaul under the floor, the quote API rejects or holds the tender for a named override. The premium cannot be negative. A fast facility is not a discount below cost. Carriers that also run an annual bid should keep this premium out of the package average. A multi-lane RFQ already has a floor on every lane. Detention risk is a note on the lane, not a subsidy.

Scorecard fields a conversation can use

A scorecard is a one-page brief a pricing lead or a planner can put on the table with the other party, still without naming them in this framework. It describes one gate over a stated window. It shows whether the clocks are good enough to talk about, and it shows the band. It does not show a demand for money. Collection remains a contract conversation using actual dwell. The scorecard's job is to make the pattern visible before the next quote, and to show the defects that should stop anyone from treating the pattern as settled.

If the valid-clock share is poor, the meeting is about the fence and the free-time field, not about a probability. If the share is sound, the meeting can be about hour of day, live versus drop, and whether the appointment bank is the stretch. Bring the band and the sample size. Leave the model version in the appendix so a later argument can be tied to a file, not to a screenshot someone remembers differently.

Cold-start facilities

A new gate, a moved pin, or a sample below the carrier's minimum is a cold start. An illustrative minimum is 30 completed stops with valid yard dwell and, where a detention probability is shown, a stored free-time rule. Below that, the scorecard says "insufficient history" in the title, not in a tooltip. The band, if shown at all, is the archetype prior and is labeled "grocery DC prior" or "mill prior" or "port ramp prior," not "this gate." The prior is wide on purpose. A tight band on five stops is false precision.

Cold-start output does not auto-apply a premium and does not auto-change a schedule. A person may still add buffer because the archetype is known to be awkward, and they write that reason themselves. When the gate crosses the minimum, switch to the gate's own band and keep the archetype in a comparison column for one quarter so a sudden jump is visible. A pin change or a new polygon resets the count. Old frontage-road hits do not carry forward into the new sample. The readiness guide covers how to tell a polygon hit from a radius hit before this model is allowed to train.

Data defects that invalidate a score

Some defects are annoying and some make the score false. The table is the second kind. If any row applies to the facility in the window you are about to quote, suppress the probability, suppress the premium, and show the defect on the scorecard. A shaded score that still looks official will be used. Suppression means the number is absent, not gray.

Fixing the defect is an operations and data task:

Retraining on the broken label does not repair it. The model will learn the defect fluently.

Human use: pricing and scheduling, not a refusal

Two roles may read a valid score, and neither role is an automatic gate. Pricing uses the probability and the excess minutes to decide whether to show the premium on a spot bid, to leave it off, or to decline the freight for ordinary commercial reasons. The decline, if it happens, is the person's. The quote API does not gain a new reject reason called "detention risk." The only hard reject in the quoting path remains the margin floor and its fail-closed rules. Scheduling uses the upper end of the band to add buffer, to prefer a drop trailer when the pool can support it, or to warn the driver that the next appointment is tight. The assignment is still made by dispatch under the carrier's current rules.

Do not wire the score to an auto-refuse on the tender, an auto-hold on the driver, or an auto-generated detention invoice. A high probability is a briefing. It can be wrong because the fence was wrong, because free time was stale, or because the gate changed last week. A person who overrides the briefing should record why, in a sentence, so the next review can tell judgment from habit. Related dispatch constraints, when a load is already accepted, belong to the dispatch constraint guide. That guide does not get to cancel a tender because a dwell band was wide.

Frequently asked questions

These answers are the boundary of the score. A demo that bills from the probability, or that refuses tenders from it, is outside this framework.

Archetype language stays on the scorecard even in the answers. No customer is implied by a grocery DC, a mill, or a port ramp.

Published October 23, 2026 by Strong Tower Consulting, LLC. Founded 2008 in Minnesota by Scott Elliott.