Machine Learning Models for Motor Carrier Driver Turnover & Churn

Who this framework 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 reader is a driver manager, a safety leader, or the person who owns the data extract. The question is narrow: can dispatch, payroll, and hours-of-service aggregates already in the company point to a conversation before a truck sits empty? Recruiting cost and uncovered freight are the business reason. The method does not reach into a driver's personal life to get there.

Why exit interviews are too late

An exit interview is a record of a decision that has already been made. The driver has given notice or has stopped showing up. The tractor is uncovered. The lane that person knew is back in the open board. Whatever the interview captures is useful for a later recruiting pitch and almost useless for keeping that seat filled this week. People also give a short, polite reason. Pay, home time, and a bad week of layover get flattened into one sentence, or they do not appear at all.

A leading review uses the weeks before anyone resigns. An illustrative label horizon is 60 days after the week being scored. The model ranks active drivers for that week. A driver manager uses the rank to decide whom to call. The call is the intervention. Waiting for the interview means the carrier studies the loss instead of interrupting it. Counting annual turnover after the fact still matters for finance. It is a lagging total, the same way a CSA score is a lagging total, and it does not name the week the plan went wrong.

This is also why a model trained only on completed exit forms learns the language of people who already left. That language is not available for the drivers who are still here. The features have to be fields the company wrote down during ordinary operations, on a clock that ended before the label window opened.

Labels, voluntary quits, and leakage

Keep the labels apart. A voluntary quit inside the horizon is the positive class for a retention model. A termination is a different event: the company ended the job, often after a safety or conduct process that has its own file. Pooling quits and terminations teaches the model to treat a firing like a resignation. A layoff, a documented medical leave, and a death in service are not voluntary quits either. Exclude them or give them their own label, and do not let them count as successes or failures of the retention list. The training row is one driver-week. The label is whether a voluntary quit is recorded in the next horizon, using only facts known at the end of that week.

The quit date is leakage if it is used as a feature. So is the separation code, the exit-interview text, the last day worked, and any inactive flag that HR sets on the way out. The defect is usually a join, not a column someone meant to cheat with. An extract that attaches today's employment status to every historical week will mark all prior weeks of a driver who later quit. The model then "predicts" an ending it was handed. Accuracy on that extract is not skill. The feature cutoff is the carrier's week ending, before the label window opens. Payroll fields are the check issued for that week, not a later void or a separation adjustment that mentions the quit.

Build the extract on a read replica, with the cutoff written into the query, the way a SQL data mining review would define any other label. Store the query with the model version. If a later analyst adds "status as of today" to make the file easier, the promotion test below will not save you unless someone re-checks the cutoff. Put that check on the calendar.

Early indicators from operations

Four patterns show up in data the carrier already posts. Unwanted layover hours are hours the driver sat that dispatch did not plan as home time and that the driver did not accept as a paid layover. Weekend dispatch frequency is the share of loads with a pickup or a delivery on Saturday or Sunday over a rolling window, compared with that driver's own earlier baseline. An illustrative window is eight weeks. Home-time misses are weeks when the promised home-time window on the dispatch plan was missed, measured as plan versus arrival at the domicile terminal or the agreed drop yard. That is an operational appointment. It is not a trace of where the driver went off duty.

Paycheck volatility from accessorial timing is the other early indicator, and it is often misread as a pay-rate problem. Detention and layover pay can be fair in a monthly average and still miss the week the driver expected them. The feature is the spread of weekly settlement net over the same illustrative eight weeks, plus the lag in days between the delivered accessorial and the check that paid it. A short week caused by a late key in billing is a process defect. It is not evidence the driver is disloyal, and it is not a reason to dock the check. The settlement audit on the driver settlement deduction page is the place to fix a ledger that drifts. This page only asks whether the timing swing is visible early enough for a manager to explain it.

Hours-of-service aggregates already retained for compliance can sit beside those fields: weekly driving and on-duty totals, and how often the plan consumed the week. They are totals, not a location history of off-duty time, and they are not a tool for changing a log. How those aggregates are stored is covered in the ELD and telematics API guide. Personal phones, personal social accounts, in-cab audio of private calls, and any off-duty location beyond the position record safety compliance already requires are out of this model.

Feature table and the week-end cutoff

Every feature in the table has to be knowable at week end from dispatch, payroll, or an hours-of-service aggregate. If a clerk types it after the driver leaves, it is not a feature. The rows are an illustrative dictionary for a pilot, not a vendor schema and not a claim that any fleet has measured these weights. Tenure in months is included because it is an employment fact. Age is not tenure. Equipment type and team or solo are included because they change how the plan feels. Name, photo, and home ZIP code are not on the list.

A simple rule should be scored next to the model before anyone treats the score as an improvement. An illustrative rule is: flag a conversation when home-time misses and weekend share are both above the level that driver's own prior quarter showed. If the model cannot beat that rule on a later block of weeks, do not deploy it. Business process optimization is the work of deciding which of these columns a manager can actually act on. A feature nobody will discuss in the conversation does not belong in the file.

The intervention is a conversation

The output is a list for the driver manager: driver id, week, a band the carrier defines, and the two or three features that moved the row. An illustrative set of bands is routine, watch, and review. Review means a conversation that week. The subject is home time, layover, or why a detention line missed the check. A model flags. A person decides. There is no automated write-up, no points posted to a personnel file by the scorer, and no portal message that tells the driver a model expects them to quit.

The conversation has to be allowed to change the operation, not only the driver's file. If the mover is a late detention key, billing fixes the timing. If the mover is weekend freight the driver did not agree to take, dispatch changes the next plan. Writing a coaching form while the plan stays the same teaches people that the score is a trap. The manager records that a conversation happened and what the company will change. The manager does not record a predicted quit date. ELD and EDI integration work belongs in the pipeline that feeds the aggregates. It does not belong in the sentence the manager says to the driver.

Fairness and protected characteristics

Do not put protected characteristics in the feature store. That includes age, sex, race, color, national origin, religion, disability, marital status, pregnancy, and veteran status. Do not add proxies assembled to stand in for them: home ZIP code, a photograph, a name parsed for origin, or a language preference. A model that scores higher on a test file after one of those fields is added is a model you do not ship. Tenure, equipment, team or solo, and the operational counts above are the allowed neighborhood of features.

Error rates still have to be looked at. If the review band flags one group far more often than another, the response is to fix the features and the threshold, not to hide the table. Use groupings HR already maintains for equal-employment reporting, inside HR's process. Do not copy those fields into the feature store to "control" the model. This page is an operating framework, not legal advice. Counsel should read the feature list before the first pilot, and again if someone proposes a new column. A fractional freight CIO review is a practical way to keep the feature list, the cutoff, and the access rules in one place.

Time-based evaluation

Do not random-split drivers or weeks. A random split lets later weeks of the same person, and sometimes the same fleet's season, leak into training. Hold out the most recent block. An illustrative holdout is the latest eight or twelve weeks, scored only with features that were knowable at each week end. Report how many people in the review band voluntarily quit inside the horizon, and how that rate compares with the fleet's base quit rate in the same block. Also report how many quits the list missed. Missing some quits is acceptable. A list longer than managers can call is not.

Do not treat a single accuracy number as a promotion test. Publish the comparison against the simple home-time and weekend rule on that same holdout. If the model does not improve the conversations a manager can finish, keep the rule. An illustrative capacity limit, not a measured result, is a review band of about the top 15 percent of active drivers in a week, and only if that count fits the managers on the board. Retrain on a schedule the carrier writes down. A fit that used a quit date, even once, is thrown out. The same time-based habit is described in the AI readiness evaluation guide.

What a high score must not do

A high score must not cut miles. It must not pull a driver off a good lane to test whether they stay. It must not change pay, block promised home time, or schedule a surprise ride-along framed as punishment. Those moves can manufacture the quit the model named, and they are unfair to someone who has not resigned. Dispatch keeps offering freight under the same rules used for every other driver. Miles offered versus miles run may be a feature. It is not a lever the scorer is allowed to pull.

The score is also the wrong input for who receives a newer tractor, a dedicated account, or a team assignment. Those choices need their own written criteria and a person accountable for them. Feeding the retention score into those choices turns a conversation list into a silent penalty. The table below is illustrative policy. It is not a discipline ladder and not a pay table.

Privacy of the score

The score sits in a restricted store. Driver managers see the people on their own board. Executives see counts by terminal unless they are part of a specific conversation. Analysts see the columns their role requires, on the replica, with access logged. The driver is not shown a raw probability on a settlement portal or a mobile app. If the company tells drivers that a retention review exists, the notice says a manager may look at operational patterns such as home-time misses, weekend freight, and pay timing. It does not claim the company scores anyone's private life.

Keep the weekly row for a defined period and then delete it or collapse it to counts. An illustrative retention period is 24 months. The score is not a job reference, not a field on an insurance application, and not a file sent to a prior-employer inquiry. A departing driver's score is closed with the other restricted HR extracts. Data protection and IT strategy planning are where the access review and the deletion job are written down. The cybersecurity guide at trucking cybersecurity covers the broader control set around that store.

Operating cadence

Run the same weekly loop every week. Neither fix is an automatic personnel action.

Once a quarter, confirm that no feature is filled in after the week ends, that quit dates remain on the label side only, and that the review band still fits manager capacity. Retire a column that never comes up in a real conversation. The quarterly check is also the moment to confirm the program still uses only dispatch, payroll, and compliance aggregates. If a vendor demo offers phone metadata or off-duty location as a "lift," that feed stays out. The readiness questions for that refusal are the same ones in the readiness guide.

Frequently asked questions

These answers restate the operating rules above. They are not legal advice, not a pay policy, and not a promise that a model will reduce turnover in any particular fleet.

Thresholds mentioned here, including a 60-day label horizon, an eight-week window, a top slice of drivers, and a 24-month keep period, are illustrative policies a carrier can adopt or replace. They are not measured results.

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