Back to all
A deep dive into the machine learning models behind more accurate ETA predictions
TL;DR
- Many retailers still build delivery promises by combining a processing window, a carrier transit estimate, and a buffer.
- That approach starts to break down in supplier-fulfilled ecommerce, where processing times vary by warehouse, backlog, fill rate, and current operating conditions.
- Rithum’s Delivery Promise uses machine learning to predict processing time and transit time separately, which produces a more realistic ETA.
- The advantage comes from the data behind the prediction, especially supplier-warehouse visibility across the network.
- Package Predictor is separate from Delivery Promise, but it improves shipping-cost decisions by predicting package weight and dimensions more accurately.
For a lot of retailers, delivery promise still starts with simple math: three days to process, five days in transit (call it eight) and move on. That kind of estimate can work for a while, especially when the fulfillment network is relatively predictable. But it gets shaky fast in supplier-fulfilled ecommerce, where one order may ship from a warehouse running normally, while the next may come from a location dealing with backlog, fill rate issues, or a completely different operating rhythm.
Carrier performance adds another layer of variability, changing by service level, lane, and time of year. So, while a static estimate may look clean in the system, it can end up far removed from how an order is actually likely to travel.
That is the problem that machine learning is helping solve in Rithum’s Delivery Promise. Instead of relying on one rule that tries to cover everything, Delivery Promise can use historical and real-time data to make a better estimate of how a given order will progress through fulfillment and transit. And because Rithum sits across a broad supplier network, the model can work from a fuller picture than most retailers have on their own.
Why do static ETA estimates break down in supplier-fulfilled ecommerce networks?
Static promise logic assumes fulfillment behaves like a fixed process. In supplier-fulfilled ecommerce, it rarely does.
If you are relying on a standard processing window and an average transit estimate, you are treating every order like it moves the same way when it doesn’t. Supplier performance is rarely consistent across the network; one warehouse may be operating normally while another is slowed by backlog, labor constraints, fill rate issues, or the type of orders coming through.
You may not know ahead of time which warehouse will handle the order. That alone makes it tough to pin processing time to one standard number.
A simple estimate is easy to put in place. Keeping it useful is another story once the network gets bigger and more complex.
How does machine learning improve delivery promise accuracy?
Rithum’s approach starts by separating two questions that many systems treat as one.
Delivery Promise uses predictive machine learning models to predict how long an order is likely to take to process before it ships and how long it is likely to take in transit after it leaves the warehouse. Those two steps are connected, but they are not driven by the same conditions.
Processing time depends on what is happening inside the supplier’s operation. Transit time depends on what happens once the package is in the carrier network. Treating them separately gives you a more realistic ETA than rolling everything into one estimate.
Rather than forcing every shipment through the same assumption, the system can use historical performance and current conditions to make a better prediction for the specific order in front of it.
Why does Rithum’s network give the model a clearer view of ETA risk?
The model only gets you part of the way. ETA accuracy also depends on how much of the fulfillment picture the system can actually see.
In supplier-fulfilled commerce, retailers are often working with gaps. They may know the supplier, but not the warehouse that will ship the order. Even when they know the likely location, they may not have a current view of backlog, fill rate, or how that warehouse has been performing under similar conditions.
Rithum works from a broader set of signals across its network, including where inventory sits, which warehouse is likely to fulfill the order, how that location has performed in similar situations, and what current conditions look like in real time.
That broader view is the real advantage. A retailer may know its own order history. Rithum can pair that with network-level visibility into supplier warehouses, which gives the model a stronger read on where risk is building and where a promise is more likely to hold.
Why do more accurate delivery promises help at checkout?
At checkout, the estimate has to hold up. When the date is built from a simple estimate, retailers usually have to play it one of two ways: pad it to be safe, or tighten it and hope the order moves the way the system expects. Neither is a great option.
With a better prediction behind it, the system can generate a date based on how that order is likely to move through fulfillment and transit under current conditions, rather than applying one broad assumption across the board.
That gives retailers a better shot at posting a date that can hold up without pushing it farther out than necessary—protecting checkout conversion rates while safeguarding brand trust.
What is Package Predictor, and how does it connect to Delivery Promise?
Package Predictor is related to Delivery Promise, but it is not doing the same job.
Delivery Promise is trying to predict timing. Package Predictor is trying to predict how the shipment will actually be packaged, especially when it comes to weight and dimensions.
That is a different problem, and it affects a different part of the shipping decision. The size of the box usually is not what determines how fast something moves through the network, but it does affect shipping cost and service selection.
That is where things get messy in dropship. You may have catalog data for an item, but not enough detail to know how a real order will be packaged, especially when multiple items ship together. And when that data is coming from a broad supplier base, it’s often incomplete, inconsistent, or both.
Package Predictor gives the system a better way to work through that uncertainty. It looks at historical shipment behavior and uses those patterns to make a better estimate than a manual default can.
How does Package Predictor improve shipping-cost decisions?
Package Predictor gives the rate estimate better information to work from. If the estimated weight and dimensions are wrong, the estimated shipping cost is wrong. And once the cost estimate is off, it becomes much easier to choose the wrong service or make a fulfillment decision that costs more than expected.
Package Predictor improves both accuracy and coverage, reducing the number of cases where the system has to fall back to broad supplier-level or retailer-level defaults.
Better package estimates sharpen carrier-rate accuracy—and that’s what drives smarter shipping decisions.
Why rising shipping complexity puts more pressure on ETA accuracy and package data
Shipping has gotten more expensive in more complicated ways. Carrier agreements are more layered than they used to be, dimensional-weight charges continue to hit certain shipments harder, and small inaccuracies in the data can create bigger downstream problems than they once did.
That puts Delivery Promise and Package Predictor under a brighter light. One is trying to generate a delivery date that holds up in a more variable network. The other is trying to improve the package data behind the rate estimate, so the shipping decision is built on something more reliable than a rough default.
When costs tighten and variability increases, the quality of those inputs is what ultimately protects your profit margins.
What retailers should do next if static delivery estimates are starting to fall short
The old approach gets harder to trust as fulfillment spreads across more suppliers, more warehouses, and more moving parts.
If you are trying to improve ETA accuracy in supplier-fulfilled ecommerce, broad averages only get you so far. Better predictions come from data that reflects how fulfillment and transit are actually performing.
Machine learning becomes useful when it can model that day-to-day variation instead of smoothing it over with one broad rule. That is especially true in supplier networks, where the operating conditions behind an order can change from one warehouse to the next.
The prediction is only as strong as the visibility behind it. The clearer the view into supplier warehouses, fulfillment conditions, and transit performance, the stronger the delivery promise becomes.
To learn more about how Rithum supports delivery promise accuracy and shipping-cost decisions, schedule a demo for a closer look at Delivery Promise and Package Predictor.
Talk to our teamKyle Knoblock is Staff Product Manager, Retailers at Rithum.
