Dynamic Rate Intelligence & Machine Learning Bidding

Who this framework is for

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

Why a lane average rate misleads

A lane average answers the wrong question. It reports what a mixed pile paid per mile. The bidder needs the slope: how accepted linehaul moves when the conditions of this tender change. Five mixes hide that slope.

Contracted primary freight and spot recovery freight can share a city pair and still be different populations. Averaging them produces a rate nobody was paid. Headhaul and backhaul hide the same way. A backhaul accepted to reposition the truck can print a low rate per mile. Copy it onto the loaded direction and the round trip loses money. The row needs a direction flag before it enters a training set.

Lead time changes what shippers will accept. Four hours of notice is not five days of notice. Equipment changes the cost floor: dry van, reefer, and flatbed do not share one elasticity. Seasonality, stored as week of year, moves reefer produce, van retail peaks, and flat construction on different calendars. One annual average washes those movements into a calm number. That calm number is what makes a bidder late.

Lane elasticity and the regression baseline

Lane elasticity, in executive language, is how much the accepted linehaul rate per mile moves when one condition changes and the others are held roughly still. Miles stay in a band. Equipment stays the same. Direction stays the same. Lead time, day of week, or week of year moves. The change in accepted linehaul per mile is the slope a pricing manager can explain to finance.

Teaching illustration only, not a market quote and not a figure from any carrier's book: accepted linehaul on a lane about 8 cents per mile higher when lead time falls from 72 hours to 12 hours. A lane average cannot show that slope. A regression can.

Start with a transparent model. Regularized linear regression, ridge or lasso, is the first control when coefficients need to be readable in cents per mile. Gradient-boosted trees are the first control when interactions matter, provided feature importance is published with every model version.

A neural bidder is a later experiment, not the first control. It belongs after three things already work: the margin floor is enforced outside the model, the fuel peg is a contract rule applied after the forecast, and a time-based holdout shows no systematic underpricing. Until those controls exist, a neural net adds a number nobody can appeal.

Feature set and future leakage

Fair at tender time, and only then: practical miles from the bid-time mileage engine (routed miles, not air miles); origin market and destination market; equipment as offered (van, reefer, or flat), not the trailer dispatched later; lead time in hours from tender to the pickup appointment; day of week; week of year; prior accept or reject history, never this tender's outcome; stop count as tendered; live load versus drop; team versus solo. Headhaul or backhaul is a required direction flag. If dispatch later changes miles, that is a new rating event.

Store the fuel peg for audit and keep it out of the linehaul matrix. The contract table applies it afterward. Any field written after the bid is leakage. Using the final invoice amount as a feature is a defect. The same defect includes a bill-to site corrected after delivery, lead time taken from actual arrival, a stop added at the dock, and a later switch to team. Lumper and detention are not features and not the target.

True operating cost and the margin floor

The floor is a policy owned by finance. It is not a model output and not a percentile of last quarter's spot loads. Finance builds operating cost per mile from the carrier's books, then states the floor in cents per mile and as a percent of linehaul revenue. The regression does not vote.

The build-up below is a teaching case for a solo dry van. Each line is an illustrative planning figure. It is not a market index, not a live fuel quote, and not a pay recommendation.

On this illustration the floor is 15 percent of linehaul revenue. Cost is $1.70, so the floor rate is $1.70 divided by 0.85, which is $2.00 per mile. The margin inside that rate is $0.30 per mile, and $0.30 is 15 percent of $2.00. Both statements are the same policy: 30 cents per mile, and 15 percent of floor linehaul. If finance changes insurance or driver pay, the dollar floor moves. Update the floor file. Do not refit the model to chase the change. Keep that file where the ledger already lives. A second cost, disconnected from accounting, will drift. Transportation accounting integration is what makes the floor defensible at month end.

Worked illustrative bid on a 480-mile van lane

Use a generic 480 practical-mile dry van headhaul. No shipper is named. Figures are round so a controller can recompute them. They are illustrative planning figures, not a rate for any real lane.

Trip cost is 480 times $1.70, which is $816. Floor linehaul is 480 times $2.00, which is $960. Dollar margin at the floor is $144. Anything transmitted under $960 linehaul is under the floor, even when it still sits above the $816 cost. The model may then suggest a linehaul. The three rows are separate illustrations, not three bids on one load.

The third row is the control working. The $1.88 suggestion clears operating cost and fails the floor. A confidence interval does not waive that stop.

Fuel peg tables stay outside the model

Carrier contracts usually adjust fuel from the EIA weekly retail on-highway diesel price. Desks often call that figure the DOE fuel peg. The peg is a published series. The surcharge is a lookup, not a forecast. A peg table has a base price, an increment, and a surcharge in cents per mile or as a percent of linehaul. Either form is allowed. Mixing them inside the regression is not.

Illustration only: base $3.60 per gallon, an increment of 6 cents, and a surcharge of 1 cent per mile for each full increment. Apply that to the auto-band suggestion above. A weekly peg of $3.90 is 30 cents above the $3.60 base, which is five increments. Five increments times 1 cent is 5 cents per mile. On 480 miles the surcharge is $24. Linehaul stays $1,080. Linehaul plus this surcharge is $1,104, before any other accessorial. The model target remains $2.25 per mile of linehaul, not $2.30.

The model forecasts linehaul. The peg adjustment is a deterministic contract rule applied after the model. The model is not allowed to learn the peg away. If training uses all-in cash, the fit absorbs yesterday's surcharge into the linehaul. The TMS then adds today's surcharge on top, and the carrier double-counts or, after a peg drop, under-recovers. Store the peg on the tender so audit can replay the table. Leave it out of the feature matrix. Fuel tax is a different problem. An IFTA mileage audit does not set the bid and must not be folded into the rate target.

Bid-band governance

Distance is measured from the floor, not from the model's favorite rate. No one, including an API, may transmit a rate under the floor without a named executive override that is audited. Model confidence is not a name.

Give the auto band a ceiling the carrier chooses. A suggestion far above the floor can be a unit error or a real tight market. The ceiling forces a look. Log the floor-file version next to the model version. Put the same bands in front of any tariff engine. A rate matrix can propose a cell. The floor still wins. See the rate matrix and tariff engine guide for the engine, and this page for the stop in front of it.

Training data hygiene

The target is linehaul revenue per practical mile, not all-in cash. Lumper, detention, layover, and driver assist are accessorials. They do not belong in the number the model is learning to bid. If the source system stores one cash amount, the row is unfit until accounting separates linehaul from accessorials.

Exclude rejected tenders from the rate regression, or model them as a separate win-probability problem. A rejected tender has no accepted linehaul. Training on the rate that lost teaches a price at which the carrier failed, or it teaches noise. A win-probability classifier can use the offered rate and the accept or reject flag. It does not replace the floor, and its output is not a rate.

Do not train a headhaul rate on backhaul loads mixed in without a direction flag. If the flag is missing, drop the row. Guessing direction from which end is closer to a terminal is how backhaul economics leak into headhaul bids.

Winsorize or review detention-bloated invoices so accessorials do not inflate the linehaul target. Illustrative review rule: if detention plus lumper is more than 25 percent of the invoice, a person confirms the linehaul split before the row enters the extract. That review is not a quiet trim that manufactures margin. A real linehaul outlier, such as a permitted over-dimensional move, should be segmented, not clipped into the van average. Keep the extract in version control. SQL data mining is the place to define accepted linehaul, practical miles, and these filters.

Cold-start lanes and uncertainty

New lanes will not have a stable slope. Pool history in this order: the exact lane, then the market pair, then the region pair, then a mileage band with the same equipment and the same direction. An example band is 400 to 500 practical miles, van, headhaul. Stop at the first pool that meets the carrier's minimum row count.

The minimum is a policy. An illustrative threshold is 30 accepted linehaul rows. Below that, step up the hierarchy and widen the uncertainty band. Show the pool's central linehaul per mile and the band. Do not print a false-precise rate such as $2.137 on a thin lane. If the half-width of the band is larger than the auto-accept window, the reason code is cold-start review and the API stays silent. Region and mileage-band pools are context for a person. They are not permission to auto-bid.

Evaluation that protects margin

R-squared is not an acceptance test. A model can fit a large share of history and still sit a few cents under the rates that protect the floor. Judge the holdout in cents and in money retained.

Hold out the most recent 8 weeks as one contiguous block. Do not random-split a time series. A random split lets later seasonality leak into training. On that block, report mean absolute error in cents per mile, and fail a promotion if the MAE is wider than the auto band the desk intends to trust. Report bias as the mean of suggested linehaul minus accepted linehaul. Negative bias is systematic underpricing and is worse than the same sized symmetric error. An illustrative teaching fail, not a measured result, is bias worse than 3 cents per mile negative. Report win rate on tenders actually transmitted, and fail a rise in wins that was bought under the floor. Report realized margin versus the floor on accepted loads, and fail any under-floor accept that lacks an audited executive override. R-squared may be filed with the version. It is not a passing grade. Compare a candidate with the prior version on that same holdout before anyone promotes it.

Operating placement

Run the scorer against a read replica of historical tenders. It does not lock the live tender table, and it does not open a session to a shipper, a portal, or a load board. It writes a suggested linehaul rate per mile and a reason code, such as within band, near floor review, under floor reject, or cold-start wide band. It also writes the model version, the floor version, and the pool level. It does not write a transmitted rate.

The TMS rating clerk or the pricing API applies the floor. That step rereads the suggestion, recomputes the floor from the current cost file, and applies the band table. Under the floor, the transmit path is closed. This is advisory architecture, not an autonomous bidding product. A model may recommend a rate. A margin floor and a person stop a bid that is below true operating cost. The same split works in front of a carrier rate matrix. Dispatch constraint solving, covered in the predictive dispatch guide, is a different decision and must not silently rewrite a rate that already cleared the floor.

Weekly operating cadence

On Monday, finance confirms cost inputs that moved: driver pay, the miles-per-gallon assumption, insurance, and tolls. Those updates change the floor file. They do not, by themselves, justify a new model fit. When the weekly EIA diesel figure is published, update the peg lookup only. Do not refit linehaul because the peg moved.

Through the week the replica scores tenders. Clerks clear the auto band. Managers clear the near-floor band. Under-floor rows wait for a named executive or they expire. On Friday, compare realized linehaul margin on accepted loads with the floor that was in force. A small MAE with negative bias is the alert. Once a quarter, confirm that every feature is still knowable at bid time. A field dispatch started keying after delivery has become leakage. That review belongs with IT strategy, process, and, when one person must hold the floor, the peg table, and the rating path together, a fractional freight CIO review.

Frequently asked questions

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