Our client operates one of the largest air cargo networks in North America, serving hundreds of international and domestic lanes. Their commercial team quotes spot rates to freight forwarders for shipments ranging from small parcels to multi-tonne pallets, across routes with wildly different demand dynamics, competitive intensity, and capacity profiles.

The core pricing formula was sound. The problem was that its main dynamic lever — the pricing adjustment factor calibrating rates based on market conditions — was determined by looking up a value in a static grid rather than computed from live signals. This grid was maintained manually and had no mechanism to incorporate feedback from actual pricing outcomes. The client needed a system that could learn from the market and respond in real time, without sacrificing the transparency and control their commercial team relied on.

One-Size-Fits-All Rate Setting

The pricing engine relied on static lookup tables that returned a pricing adjustment based on a limited set of inputs — current capacity utilization and time to departure. Every quote that fell into the same cell received the same adjustment, regardless of the lane's market share, competitive dynamics, or how that specific flight's bookings compared to its historical trajectory at that point in the booking window.

Key-Person Fragility

The logic behind the table values — why specific cells were set to particular discount levels and when they should be updated — resided entirely in the institutional knowledge of experienced pricing managers. There was no formal feedback loop from pricing outcomes back to table values and no version-controlled record of reasoning behind past changes, leaving the tables increasingly detached from market reality over time.

Revenue Leakage at Scale

Air cargo booking patterns are heavily back-loaded, with a disproportionate share of volume booking close to departure. While the static tables factored in current capacity utilization, they responded only to the absolute fill level — not whether that level was ahead of or behind the route's historical norm at that point in the booking horizon. A flight running ahead of expectations was priced identically to one that was lagging, missing the opportunity to capture yield in the most critical window.

Compass built a three-component optimization pipeline within Dataiku that transforms cargo spot pricing from a static lookup into a data-driven revenue engine. The system preserves the existing pricing formula structure and produces the same output format — ensuring zero disruption to downstream systems — but derives each pricing decision from an explicit revenue-maximization objective instead of a fixed grid.

The first component is a capacity adjustment that measures how far a flight's current bookings deviate from its expected trajectory at a given point before departure. Historical booking curves at multiple levels of granularity provide the benchmark, with the most route-specific curve selected at scoring time. A commercial strategy parameter scales this benchmark based on each route's priority — yield-focused segments push prices up more aggressively, while growth segments favor competitive positioning. The resulting capacity signal feeds directly into the optimizer.

The second component is a win probability model that estimates the likelihood a customer accepts a quote at a given price. Compass built a logistic regression that predicts win probability based on price competitiveness, time to departure, and route characteristics. A synthetic elasticity correction sharpens the model's sensitivity across the full pricing range, extending its reliability beyond the narrow band of prices observed historically.

The third component is a per-quote constrained optimizer that selects the pricing parameter maximizing expected net revenue, subject to floor rates, minimum win probability thresholds, and segment-specific pricing bounds. The optimizer runs in real time via the Dataiku API Node. All model artifacts are retrained through automated Dataiku scenarios, every call is logged, and the legacy static output is returned alongside each result for direct comparison — enabling a controlled, evidence-based transition.

Compass replaced a rigid, manually maintained pricing grid with a self-improving optimization engine that learns from market outcomes, responds to real-time capacity dynamics, and protects commercial objectives through explicit constraints — all while preserving full interpretability and zero-disruption integration with existing downstream systems.

Win Probability from Sparse Outcomes

Win rates in competitive air cargo spot markets are inherently low, meaning any individual route accumulates very few usable outcomes over time. Compass built a logistic regression that estimates win probability based on price competitiveness, time to departure, and route characteristics. A synthetic elasticity correction ensures the model responds meaningfully across the full pricing range, and route-level pooling allows the model to learn from similar lanes rather than relying on any single route's limited history.

Benefit: The optimizer gains a reliable signal of customer price sensitivity even on routes with sparse historical data, enabling the system to price competitively across the network rather than defaulting to the conservative static table.

Capacity-Aware Price Adjustment

The static tables applied the same pricing factor to all flights at the same capacity utilization and days-to-departure combination, regardless of whether a flight was running ahead of or behind its own history. Compass replaced this with a booking curve deviation approach that measures relative capacity scarcity at multiple levels of granularity. A commercial strategy parameter scales the benchmark based on each segment's priority, so two flights at the same fill level with the same days remaining are priced very differently depending on how each compares to its own adjusted historical norm.

Benefit: The airline can now capture additional yield on flights filling faster than expected while staying competitive on flights that are lagging — a distinction the static tables were structurally incapable of making.

Autonomous Decision-Making with Built-In Safeguards

The optimizer runs autonomously but is bounded at every level. Floor rates ensure no quote falls below cost recovery or a defined minimum relative to the lane's commercial target. Segment-specific pricing bounds enforce commercial strategy. A minimum win-probability constraint prevents sacrificing all volume for margin. If the model lacks sufficient confidence or no feasible pricing solution exists within the defined constraints, the system degrades gracefully to the legacy approach. The phased rollout allows the commercial team to validate the optimizer's judgment before granting it full pricing authority.

Benefit: The commercial team retains full strategic control through explicit constraints while the optimizer handles per-quote computation autonomously at scale. The graceful fallback and phased rollout give the team the confidence to trust the system's judgment before full deployment — ensuring the transition from static to dynamic pricing is both smooth and reversible.