The decision is the product
Most analytics work sold into freight rail is priced as a model and consumed as a decision, and the gap between those two things is where the money goes. A railroad does not buy a probability. It sets out a car, or it does not. It slow-orders a segment, or it does not. It sends a rail-flaw crew to milepost 118 before milepost 204, or the other way around. Each choice has an owner, a cost when it is wrong in each direction, and a paper trail a regulator or a car owner can pull months later. Start from the choice and work backward to the model, and the engagement finishes. Start from the model and hunt for a decision to attach it to, and it stalls in a pilot nobody can turn off or turn on.
This is more true in rail than in most industries, because rail already standardized its decisions. The Association of American Railroads and its subsidiary Railinc run shared systems that every interchange carrier touches, with a fixed vocabulary of alerts, defects, and repair responsibilities. A score with no place in that vocabulary is a score nobody can act on. That constraint sounds limiting. It is the best thing about the domain, because the target variable is already defined by somebody other than the vendor.
The industry already picked the output format
Railinc's Equipment Health Management System is the clearest example. EHMS communicates equipment condition and sends alerts to responsible parties, generating those alerts from wayside detectors. Its alert levels are graded, and the grades carry different operating consequences: a window open status advising that some degradation has started; an AAR A2 alert indicating equipment may be repaired if it is in a shop for any other reason; an AAR A1 alert indicating equipment may be pulled into a shop specifically for that repair; and a mandatory alert telling carriers and car owners that high stresses are being placed on rails, requiring immediate action.
Read that ladder as a machine learning specification and it writes itself: four ordered classes, each with a different cost of a false positive, all keyed to equipment registered in Umler, the Railinc system holding specifications for more than two million pieces of North American rail, steamship, and highway equipment. Railinc's Asset Health Data Summaries pull from Wheel Impact Load Detectors, the Trackside Acoustic Detection System, and truck hunting detectors across the Class I railroads and short lines alike, and the stated benefit is letting operators act on potential issues before they escalate to alert-level status. The industry published the problem statement. The open work is the escalation forecast, not the alert.
Downstream of the alert sits an equally fixed structure. Damaged and Defective Car Tracking handles damaged cars under AAR Interchange Rule 107 and defective cars under Rules 1, 96, and 108, with Rule 95 covering damaged equipment that can be returned to service. Car Repair Billing settles who pays. A model that predicts a bearing failure but emits an output that cannot be routed into that chain has produced a research result, not an operating tool. We build the routing first, because the routing determines whether anybody ever sees the score.
The bars below are our read on how much of the receiving machinery already exists for six common decisions. A high number means the alert taxonomy, the definition of an exception, the person who owns the call, and the cost of being wrong are all established, so the build is mostly the model. A low number means the surrounding workflow has to be built too, which is where budgets get consumed.
How much of the receiving workflow already exists
Editorial weighting from public sources and practitioner reading — illustrative, not a measured statistic.
The clock is what makes a prediction worth money
Freight rail runs on regulated intervals, and an interval is what converts a forecast into cash. A prediction that a bearing will fail eventually is worth nothing. A prediction that it will fail before the train reaches the location where its next inspection is legally due is worth the cost of a road failure.
The brake and inspection clocks live in 49 CFR Part 232. Section 232.205 requires a Class I brake test at the initial terminal. Section 232.207 sets the Class IA test at 1,000 miles. Sections 232.209 and 232.211 cover the intermediate and trainline continuity tests. The one that matters most commercially is Section 232.213, extended haul trains: a railroad may move a designated train up to 1,500 miles between brake tests and inspections, but only if it designates the train in writing to FRA's Motive Power and Equipment Division, performs a Class I brake test at origin using a qualified mechanical inspector, performs a Part 215 freight car inspection at origin, repairs or removes every car with a Part 215 non-compliant condition, and holds the train to no more than one pick-up and one set-out en route apart from setting out defective equipment.
That paragraph is a business case. The extended haul designation is worth real money in run time and crew starts, and it is forfeited by exactly the events a defect model predicts. Ranking inbound cars by the probability of a Part 215 finding at the origin inspection is not a science project. It protects a designation with a known dollar value, and the operator can already tell you what that value is.
| Regulatory clock | Where it lives | The operating decision | What a model can move |
|---|---|---|---|
| Initial terminal brake test | 49 CFR 232.205 | Which cars go on this train | Cars set out after departure |
| 1,000-mile inspection | 49 CFR 232.207 | Where the train takes its next test | Unplanned intermediate stops |
| Extended haul designation | 49 CFR 232.213 | Whether the train keeps 1,500-mile status | Part 215 findings at origin |
| Continuous rail testing | 49 CFR 213.240 | Which suspect locations get verified first | Field-verification yield per crew day |
| Concrete-crosstie automated inspection | 49 CFR 213.234 | Once or twice a year, at the 40 mgt line | Exceptions found per pass |
| Daily locomotive inspection | 49 CFR 229.21 | Which units get pulled from the pool | Power failures en route |
Continuous rail testing is a ranking problem the regulator already specified
Section 213.240 is worth reading closely if you sell anything into track maintenance. It lets a track owner satisfy internal rail inspection requirements through continuous testing rather than stop-and-verify, and it states what the owner's procedures have to guarantee: that test data is timely and accurately transmitted and analyzed; that suspect locations are accurately identified for field verification; that suspect locations are categorized and prioritized according to their potential severity; that they are accurately field-verified; and that they are designated after verification.
That is a triage specification written by a safety regulator, and it names the metric. The number that matters is field-verification yield: of the suspect locations a crew was sent to inspect, what share turned out to be real, and how severe. A model that raises yield lets the same crew cover more track without lowering the confirmation rate. One that lowers yield burns crew hours, and the operator sees that within a week. It is one of the few places in this industry where the success criterion, the labels, and the cost of a false positive are all fixed before anyone writes code.
Track geometry is close behind, though the rules there are unsettled. Section 213.234 already requires automated inspection technology to supplement visual inspection on certain concrete-crosstie main track, twice a year above 40 million gross tons annually and once at or below it. FRA proposed in October 2024 to require Class I and Class II railroads, intercity passenger railroads, and commuter railroads to operate a qualifying Track Geometry Measurement System, and extended the comment period that December. That proposal has not been issued as a final rule. Treat the mandate as a possible tailwind, not a scheduled one.
Crew is the constraint most models ignore
Optimization vendors love crew scheduling, and crew scheduling in U.S. freight rail has hard walls written into the CFR. Subpart G of 49 CFR Part 218, Train Crew Size Safety Requirements, sets the baseline at Section 218.123(b): except as provided in the subpart, each train shall be assigned a minimum of two crewmembers. Sections 218.125 and 218.127 carry the passenger, tourist, and freight exceptions; Section 218.129 carries conditional exceptions tied to compliance dates for certain Class II and Class III operations, work trains, helper service, and lite locomotive movements; and Sections 218.131 through 218.137 set out the special approval petition, the required risk assessment content, and the railroad's annual responsibilities afterward.
Section 218.123(c) then removes the discretion where it matters most. For a high-hazard flammable train, a train carrying twenty or more loaded tank cars or intermodal portable tanks of specified hazardous materials, or a train carrying rail-security-sensitive materials as defined in 49 CFR 1580.3, the exceptions in Sections 218.125 and 218.127 do not apply at all. Any tool treating crew size as a free variable will eventually propose an assignment that is not a schedule change but a petition to the regulator. Encode the carve-outs as constraints before the optimizer runs.
Fatigue is the adjacent constraint. Subpart G of 49 CFR Part 271 requires a Fatigue Risk Management Program inside the Risk Reduction Program, and Section 271.3 applies that part to Class I railroads, to railroads FRA determines have inadequate safety performance under Section 271.13, and to volunteers under Section 271.15. Commuter and intercity passenger railroads sit under the parallel System Safety Program in Part 270. A crew-calling model that changes rest patterns has walked into a program with a filed plan and an internal assessment cycle, and compliance needs to see it before operations deploys it.
The benchmark is public, which is unusual and useful
Rail is one of the few freight modes where an outside party can check an operator's performance without asking permission. Under 49 CFR Part 1250, each Class I railroad reports performance data to the Surface Transportation Board weekly. Section 1250.2 lists what: system-average train speed overall and by train type, broken out for intermodal, grain, coal, automotive, crude oil, ethanol, and manifest service; terminal dwell in hours for the system and the ten largest terminals by railcars processed, excluding run-through cars; cars on line by car type; dwell at origin for unit trains; trains holding per day by train type and by cause, split into crew, locomotive power, and other; and loaded and empty cars not moved in 48 hours or more. Section 1250.3 adds Chicago terminal reporting and Section 1250.4 adds rail infrastructure project reporting. STB publishes both the individual carrier reports and a consolidated spreadsheet, and it keeps amending the rule — a recent final rule in Docket EP 724 (Sub-No. 5) added certain chemical and plastics traffic as a distinct category within the cars-held metric.
The safety side is just as open. FRA's public data holdings on the Department of Transportation data hub include the Rail Equipment Accident/Incident file from Form 6180.54, which as of mid-August 2026 held 224,941 records across 133 fields spanning January 1975 through May 2026 — cause codes, milepost, latitude and longitude, track class, recorded speed, gross tonnage, hazmat cars involved, hours on duty for engineers and conductors, and a free-text narrative on every record. The Highway-Rail Grade Crossing Incident file from Form 6180.57 held 251,149 records, and the current Form 71 crossing inventory held more than 438,000 records. FRA refreshes these continuously.
Two consequences follow, both commercial. A supplier can build and validate a national baseline without a single customer contract, then walk into a meeting with a result rather than a proposal. And a buyer can demand that any vendor claim be reproduced against that public file. We treat the second as the price of admission: if a claimed lift cannot be reproduced on data the customer can download, it is not a claim, it is a brochure. The mechanics are the same ones in verifying a vendor's benchmark claim.
One caution on that safety file. Whether an event is reportable at all depends on a damage threshold that moves: Section 225.19(e) requires the FRA Administrator to publish, no later than November 30 each year, a notice announcing the reporting threshold taking effect the following January 1, and Part 225 sets out the wage-and-equipment index formula behind it. A model trained on a multi-decade accident file is therefore trained on a label whose inclusion criterion drifted upward over time. Adjust for it, or the model learns the threshold instead of the physics.
Class I, short line, and intermodal are three different businesses
Buyers in this space get treated as one market and are not. The Surface Transportation Board classifies carriers by annual operating revenue under 49 CFR Part 1201, deflated to a 2019 benchmark using a producer price index for line-haul railroads. For 2025, the most recent year for which STB has calculated deflator factors, the Class I threshold was $1,094,774,354 and the Class II threshold was $49,143,204. The Association of American Railroads counts six U.S. Class I railroads. The American Short Line and Regional Railroad Association counts 603 short lines operating 47,500 route miles, twenty-nine percent of U.S. freight rail, providing service for one in five cars moving each year.
That asymmetry drives everything about a first engagement. A Class I already has the data plumbing, because Part 271 makes it file a Risk Reduction Program and Part 1250 makes it publish its own service metrics weekly. Its problem is ranking inside streams already flowing. A short line usually has the opposite problem: the detector readings exist, the work orders exist, the interchange records exist, and nothing joins them. For that operator the highest-value deliverable is frequently a joined record with clean equipment identifiers, delivered before any model is trained — useful on its own, and the thing the model cannot be built without.
| Dimension | Class I | Short line or regional | Intermodal operator |
|---|---|---|---|
| Revenue classification (2025) | Above $1,094,774,354 | Class II above $49,143,204; Class III below | Not classified by rail revenue |
| Risk Reduction Program | Required under 49 CFR 271.3 | Only on an FRA inadequate-performance finding, or voluntarily | Outside the part |
| Public performance benchmark | Weekly STB filing under 49 CFR Part 1250 | No equivalent public series | No equivalent public series |
| Federal funding path | Through partnership with a State or subdivision | Directly eligible under 49 U.S.C. 22907 | Through a partnership route |
| Where a first build pays | Ranking inside an existing alert stream | Joining records that were never joined | Reconciling interchange events |
Where the federal money is, and what it demands
The Consolidated Rail Infrastructure and Safety Improvements program at 49 U.S.C. 22907 is the main federal grant channel touching this work, and its eligibility list is friendlier to technology than most people expect. Eligible recipients include States and their political subdivisions, public agencies, Amtrak and other intercity passenger carriers, Class II and Class III railroads, associations representing those railroads, federally recognized Tribes, university transportation centers doing rail research, and — the clause suppliers miss — any rail carrier or rail equipment manufacturer in partnership with a State or a political subdivision.
Eligible projects at subsection (c) start with deployment of railroad safety technology, including positive train control and rail integrity inspection systems. The list also covers short-line and regional infrastructure, development of a safety program or institute designed to improve rail safety, research the Secretary considers necessary to advance rail capital, operations, or safety improvements, and workforce development. Selection under subsection (e) gives preference to a proposed federal share at or below fifty percent, then maximizes net benefits based on a cost-benefit analysis.
That cost-benefit language shapes the technical proposal. A reviewer is not scoring model architecture. They are scoring avoided incidents, avoided delay, and avoided damage against project cost — which is why the public FRA accident and crossing files matter, because they are the natural denominator for that arithmetic and the reviewer can check them. FRA published the combined FY2025 and FY2026 CRISI notice of funding opportunity on April 22, 2026, making up to $2,039,246,480 available, and later extended the deadline to June 25, 2026. That cycle has closed. Assemble the partnership and the benefit case before the next notice appears, because the cost-benefit exhibit is the long-lead item, not the application form.
Intermodal fails differently
Intermodal freight shares the rail network but not the failure mode. The binding document is the Uniform Intermodal Interchange and Facilities Access Agreement, administered by the Intermodal Association of North America and described by IANA as the only standard industry contract setting the rules for interchange of equipment between intermodal trucking companies and equipment providers. Roughly 13,000 trucking companies participate, and about 95 percent of North American equipment interchanges run under it.
What that standardizes is the contract, not the data. The recurring loss in intermodal is not a mispredicted arrival; it is an interchange event that two parties record differently and nobody can reconcile — a chassis whose ingate and outgate do not pair, a per diem clock that starts on one system and not another, a container physically present and administratively missing. Entity resolution and event reconciliation carry more value per engineering hour here than forecasting, and the result is auditable in a way a forecast never is. Our framing for ranked, explained exceptions rather than opaque risk scores is in freight exception handling without a black box.
Structural change is the risk nobody puts in the model
Route mixes, interchange points, and reporting marks are not stationary, and 2026 is a live example. Union Pacific's proposed control of Norfolk Southern sits before the Surface Transportation Board in Docket No. FD 36873. In Decision No. 21, effective May 28, 2026, the Board accepted for consideration the revised primary application filed on April 30, 2026, along with a related application covering the Peoria & Pekin Union Railway — then held both proceedings, including the environmental review, in abeyance pending further Board order, requiring supplemental information from the applicants by July 27, 2026. Outcome and timing are genuinely unsettled, and anyone telling a customer otherwise is guessing.
The engineering response is not to forecast the merger. It is to build the identifier and geography layer so a change in reporting marks, a moved interchange point, or a shifted route mix is a data update rather than a rebuild. Hard-coding carrier identity into features is the fastest way to own a model that silently degrades the quarter after a structural change. It costs almost nothing if it is decided at the start.
What we would build in the first eight weeks
- One joined record per car-event, with equipment identifiers reconciled against the industry registry. Useful on its own, and mandatory before any model.
- Site and season normalization on every detector reading before a single feature is computed. Raw differentials teach a model which detector it is looking at.
- A strict temporal split, never a random one. The alert that caused the set-out is not a feature; it is the label wearing a disguise.
- An output that maps onto the existing alert ladder, so the score arrives as something the shop already knows how to act on.
- A published-data baseline the customer can reproduce from the FRA files before any private-data claim is made.
- Written kill criteria in the statement of work, with the date and the number that would end the work.
- Version and snapshot lineage on every advisory that changed an inspection or a set-out decision, reconstructable months later.
Common objections we hear, and the honest answers
Our detector thresholds already work. What does a model add?
Threshold logic evaluates one reading against one number. The added information is in the trajectory of readings for one bearing across successive detectors, and in how that trajectory compares to the same car's own history and to the site's recent distribution. The industry's own asset health tooling frames the goal as acting before an alert level is reached, which is precisely a forecasting problem rather than a threshold problem.
Can any of this touch the train control system?
No, and it should not. Positive Train Control is certified under an FRA-approved safety plan per 49 CFR 236.1015, with error and malfunction handling at 236.1023 and records retention at 236.1037. Analytics reads a one-way copy and produces advisories for people. Nothing a model generates returns to the vital system.
We are a short line. Is this affordable at our scale?
The data volumes are small — detector readings and work orders are gigabytes, not petabytes — so infrastructure cost is rarely the obstacle. The real constraint is engineering time on joins and identifiers. That is also why 49 U.S.C. 22907 matters: Class II and Class III railroads and the associations representing them are directly named as eligible recipients.
Who owns the model and the data at the end?
Settle it in the statement of work, not at closeout. On federally funded work the data rights clauses govern and should be read before the kickoff meeting. On commercial work we write ownership of trained weights, derived features, and the pipeline into the agreement up front, because a model the customer cannot retrain is a rental.
Bottom line
Freight rail is a good domain for applied machine learning for an unglamorous reason: somebody else already defined the exceptions. The alert grades come from AAR, the inspection intervals from the CFR, the performance metrics from STB, and the accident labels from FRA, in public. The work that pays is fitting a defensible model into that structure and proving the lift on data the customer can download. The work that does not pay is a general-purpose platform sold into an industry that already has one.
Frequently asked questions
A large amount. FRA publishes the Form 6180.54 rail equipment accident file, the Form 6180.57 grade crossing incident file, and the Form 71 crossing inventory on the Department of Transportation data hub, with cause codes, locations, and narratives going back decades. STB publishes weekly Class I service performance data under 49 CFR Part 1250. A credible baseline can be built from those before a first meeting.
Usually ranking, not forecasting. Continuous rail testing under 49 CFR 213.240 requires suspect locations to be categorized and prioritized by potential severity before field verification, so field-verification yield is a metric the operator already tracks and can see move within weeks.
Not uniformly. Track, brake, and locomotive standards apply broadly, but the Risk Reduction Program in 49 CFR Part 271 applies to Class I railroads, to railroads FRA finds to have inadequate safety performance, and to volunteers. Crew size requirements in 49 CFR Part 218 Subpart G include conditional exceptions specific to certain Class II and Class III operations.
Yes. The CRISI program at 49 U.S.C. 22907 funds deployment of railroad safety technology including rail integrity inspection systems, short-line infrastructure, safety programs, and rail research. Class II and Class III railroads are directly eligible, as is a rail carrier or equipment manufacturer partnered with a State or political subdivision. Selection favors a federal share at or below fifty percent and turns on cost-benefit analysis.
Scope it as reconciliation rather than prediction. Interchange between motor carriers and equipment providers runs under the UIIA administered by IANA, which standardizes the contract but not the data. Most recoverable value sits in resolving equipment identity and pairing interchange events across parties, which is auditable in a way a forecast is not.