Making Machines Your Chaos Co-Pilots: Why AI Freight Rate Prediction Isn’t Like Recognizing Bicycles

The Reality Check: AI in Freight Isn’t Magic

When freight professionals think about artificial intelligence solving their rate prediction challenges, they often envision the same remarkable accuracy that AI achieves in image recognition: 97–99% precision. However, as Dr. Vishal Gupta (USC Marshall School of Business) bluntly explains in the Freightos Stable Chaos: Digital Supply Chain Summit: “The hope of developing AI tools that have the kind of accuracy we do have for vision recognition, like 97, 98, 99 accuracy, it’s just not gonna happen.”

This is not cynicism but a practical assessment grounded in real work. Gupta and Robert Khachatryan (Founder, Freight Right) have spent years trying to make rate prediction useful for operators. Their conclusion: AI adds value in freight when you start with clean data, define the right problem, and treat prediction as input to risk decisions—not as a crystal ball.

“We tend to call all of these problems rate prediction, yet… if we think about them from more of a machine learning or AI lens, they actually are kind of different.” — Dr. Vishal Gupta.

Why Freight Rate Prediction Is Fundamentally Different

Gupta illustrated the gap with a simple analogy: a two‑year‑old can label a bicycle correctly, and most two‑year‑olds will agree. That’s “ground truth.” Freight rates have no such consensus.

• “If you pick up any two‑year‑old and you ask a two‑year‑old, like, is this a bicycle? This two‑year‑old is going to be able to be like, yeah, that’s a bicycle… But if you think about now, something like predicting freight rates… I think we would get a ton of different answers. There’s not an obvious consensus.”
• “There’s also no obvious rule by which [rates are] determined. Economists have argued for a long time about whether there are market forces that determine supply and demand… or whether it truly is a random walk down Wall Street.”

Freight looks more like financial markets—chaotic, regime‑driven, and partially observable—than like computer vision.

The Data Quality Crisis

The industry’s “training data” often undermines the model from the outset, noted Khachatryan:

• “When Vishal and I started working on our project… even our data was very crappy. Because five years ago… nobody was typing any information into our systems, thinking about AI research five years down the road.”
• “We have something like 27 variations of how people… call freight.”

These aren’t edge cases—they’re the norm. A survey by Ontegos Cloud reveals that data quality issues are pervasive and costly, affecting over 65% of forwarders who face quality gaps that negatively impact customer satisfaction and lead to disputes. Furthermore, 77% lack the quality data required for AI, resulting in up to 70% of their time being spent on manual entry and corrections.

The Unsexy Solution: Data Cleaning Over Algorithmic Complexity

Gupta’s first recommendation: “Think about data cleaning.” It’s not glamorous—but it works.

Proof from Medical AI

In Google’s widely cited diabetic retinopathy work, the most significant lift came not from a bigger model—but from better labels:

• The team flagged ambiguous images and sent them to retinal specialists to create gold‑standard labels.
• With the same algorithm and cleaner data, performance jumped: “from being something that’s just below what a board‑certified ophthalmologist does, to something comparable to what a retinal specialist does.”

No extra compute—no exotic model. Better labels.

Turn Cleaning into Product Requirements

Freight Right applied the same lesson to operations, according to Khachatryan:

• “We almost… flipped from, basically, okay, you have data, you’re trying to use some kind of a mathematical model… to basically saying… everything we do now at some point is going to be analyzed in some way… we’re basically thinking AI first anytime we think about data now.”
• Gupta added that they saw an “ocean LCL quote… delivering to an inland port” with a wrong (massive) total; the fix wasn’t just to delete it. They added UX checks: if the destination is inland, require a drayage line; don’t allow blended legs.

Clean at capture. Validate with UX. Make data quality a habit, not a one‑off clean‑up.

Two Different “Rate Prediction” Problems (Don’t Mix Them)

Not all “rate prediction” is the same. Define the use case first, recommended Gupta:

• Quoting acceleration (space): For forwarders, pre‑fill realistic rates on repetitive lanes to reduce SLA from hours to minutes, cut emails, and raise quote coverage.
• Planning under uncertainty (time): For BCOs, forecast bands for tenders, budgets, and contract‑vs‑spot choices over months—delivered as scenarios, not single‑point forecasts.

“Know which problem you’re solving before you build.”

From Prediction to Risk: The Strategy Shift

Perfect forecasts are unrealistic in freight. What’s realistic—and valuable—is risk‑aware decisioning:

• “If you are honest with yourself about that reality… you’re never really going to be able to predict at that type of accuracy, the question shifts… to ‘How do I start thinking about risk management?’” — Dr. Vishal Gupta.
• In volatile markets—like the Red Sea shock that saw average 40’ prices jump from $1,339 to $3,434 in six months—treat predictions as noisy inputs to guardrails, hedges, and escalation triggers.

Think portfolio, not point estimates. Quantify uncertainty. Align actions (contract vs. spot, hedges, allocations) to ranges—not false precision.

Where AI Helps Today (With Real Impact)

• Quoting at scale: For repetitive lanes, ML can pre‑fill competitive rates and slash cycle time, with human‑in‑the‑loop review for exceptions.
• Spatial prediction: “Predicting rates in space… if you have missing rates or missing legs”—especially relevant as DTC/e‑commerce drives delivery to “all these random addresses.” — Robert Khachatryan.
  Example: If you have a last‑mile rate for 90012, interpolate a range for 90013 using distance, carrier zones, and recent deliveries.
• LCL complexity: “An average LCL [quote contains] like 27 different lines of charges. How do you predict each one of those?” asked Khachatryan. Carriers are “very protective of their data,” so ML can help classify/estimate cost stack components when tariffs aren’t exposed.
• Missing legs: Provide cost ranges for pre- and on-carriage legs to facilitate door-to-door quotes and enhance coverage.

One thing is clear: The chaos is here to stay, according to the Gupta and Khachatryan, who’ve figured out how to deliver predictable results when literally nothing else is predictable.