Machine learning is a kind of artificial intelligence. It deals with the process behind systems that use inferences and statistical models to find connections in data. Patterns in the data allow the model to draw conclusions and make predictions.

Triumph Intelligence uses machine learning models to predict prices of future loads based on patterns found in historical load data.

Artificial Intelligence is the general ability to emulate human thought and behavior through computer science, while machine learning refers to the technologies and algorithms that enable systems to identify patterns, make decisions, and improve their capabilities through experience and data. AI is an umbrella term for computer software, with many subsets that work to achieve desired human-like outcomes. Other subsets within AI include deep learning, natural language processing and robotics.

Many companies continue to explore AI as a means of increasing productivity and optimizing their workforce with automation and chatbots.

The key differences:

Machine learning is the primary way most people commonly interact with AI.

Many different types of AI models can be used to make predictions based on previous data. Different models have different structures, and each has its strengths and weaknesses.

In most cases the machine “learns” by testing examples and receiving feedback on how close each result is to being correct.

It uses feedback to adjust the weight it assigns to each factor in determining a result. Over time, the model gets better and better at predicting outcomes.

A machine learning model is a smart program that “learns” patterns and relationships from data, which lets it make predictions and perform actions without being told exactly what to do.

The algorithms and parameters that make up a machine learning model can be optimized through training for better performance and more accurate predictions.

Each sample (a load in the case of Triumph Rate Intelligence) is represented by a set of features. Features represent important information about the load, such as its origin, its destination, and the distance between them. Every historical load is also accompanied by its price, which tells the model what kind of prediction to offer, given the conditions encapsulated in the load’s features.

The AI notes and analyzes all of those variables -- where the shooter is, who the shooter is, what the defense is doing -- and finds the combination most likely to score in each situation. It can handle a huge amount of data, and it gets better as it takes more shots, so it doesn't take long for it to achieve a high shooting percentage.

Now let's apply that to freight pricing. A machine learning model can take a lot of variables (origin, destination, we`ight, fuel, etc.) into account. It notes how each variable affects rate pricing in the context of all the other variables, and predicts a rate.

A human could arrive at the same prediction, given enough time, but a machine learning system can offer reliable predictions a lot more efficiently. It can even uncover dependencies and trends that might be hard for a busy human to catch, such as time-of-day effects on capacity and price. Machine learning multiplies the speed and efficiency of the human’s work. The model, however, needs to start with real-world information–in this case, historical loads.

For years, brokerages have relied on historical averages alone to predict current freight costs. While historical data is critical in predictive models, when used alone, it can only capture past trends, while predictive AI models use both historical and current data and machine learning algorithms to predict what will happen next.

Due to the volatility of the freight market, using simple historical averages to predict current rates will often produce inaccurate and partial results. AI models use advanced statistical analysis to make predictions and measure the Confidence of those predictions. They also deal with data scarcity in a much more consistent and controlled way than plain averages. Predictive models can take into account many different features of a load, identify correlations that would be difficult to the human eye to recognize, and provide a rate that has considered all of those features.

Triumph automates the process of statistical analysis to provide reliable and easy-to-read rate predictions.

We start by dividing our historical data into three sets: a training set, a validation set, and a test set. The training set typically contains the majority of the historical data. The additional sets outside of the training set measure the model’s performance on previously unseen data.

We start by dividing our historical data into three sets: a training set, a validation set, and a test set. The training set typically contains the majority of the historical data. The additional sets outside of the training set measure the model’s performance on previously unseen data.

The training phase provides us with multiple competing models that differ in the way they operate, the features they use, or the weight they give to different parameters. For instance, one model might assign more importance to origin and distance, while another prioritizes load size. We need to select the model that provides the most accurate rate predictions, either manually or by an automatic selection workflow.

The next step uses the validation set, which contains samples previously unseen by the model, to evaluate the models and select the most accurate ones. The models are given features of the loads within the validation set without their outcomes--in this case price--and each model's predictions are compared to the actual values from the sample. Each model is graded on how closely it predicted the validation set as a whole. We choose the most accurate model and refine it further with the values from the validation set.

We still need an objective assessment of the selected model’s results on previously unseen data, so we can be sure that the success of that model didn’t involve a bit of luck. So we use the test set, which is still new to the model. This final test gives an objective evaluation of the model’s real-world performance.

The Rate Intelligence app shows the model’s performance against the test set in the Rate Accuracy Report, which measures the accuracy of the predicted rate compared to the actual booked rate for the load, so you can see how close your model is coming to actual rates at any time. If the model seems off, our engineers can take a look and find out why.

If we tested using the original training set, all we would find out was how good the model was at parroting back information it already had. Using a new set tells us how well the model does with new information.

For instance, if we trained the model with a training set that included a shipment from Boston to San Antonio, testing the model with the same set would only test whether it could repeat the same rate every time. That wouldn't be much use, so we test with a different set.

Data sources provide the information from which the model learns. In the case of Rate Intelligence, the model uses a few different data sources. The primary data set used to predict rates for your brokerage is your own historical load data.

Origin and destination, pickup date, transport type, transport mode, carrier and customer costs, and linehaul cost are the most important features for rate prediction. Rate Intelligence models also learn from a wider set of brokerage data from the Triumph Network. It can also learn from external data sources and macroeconomic features for increased accuracy.

The Verified Buy Rate is the predicted buy rate on a lane for that given brokerage.

Rate Intelligence predicts this rate when you enter a lane and transport type into our user interface or an integrated TMS. This rate is influenced by your historical data and reflects your company’s individual buying behavior against the current market conditions, but is also often influenced by external data, such as the data we receive for other brokerages (the Triumph Network).

The Verified Sell Rate is determined as a markup of the Verified Buy Rate. The markup value can be a percentage or a dollar value.

It can be determined through the creation of pricing rules, or generated by the customer’s historic margin on the lane. Without rules or historical data, the markup will default to a baseline of 15%. Here’s an example of a Verified Buy Rate: let’s say that for a particular shipment we predict a rate of $1,000. Currently the brokerage has a pricing rule stating that all loads originating in Seattle should have a 10% markup. The sell rate in this case would be $1,100. A markup of 20% would return a rate of $1,200, and so on.

The Market Rate represents the expected rate in the market based on the aggregated data in the Triumph Network.

The Market Rate has no bias towards any one brokerage’s data. It can be lower or higher than the Verified Buy Rate, depending on how the brokerage books freight on the underlying lane or similar lanes.

So how does Rate Intelligence handle lanes that don’t see much traffic? Maybe no one has historical data on a shipment from Butte, MT to Eureka Springs, AR, but he machine learning model can still come up with an accurate prediction on the lane, because there’s plenty of data available for shipments from Spokane to Little Rock that use the same corridor.

A Rate Intelligence machine learning model recognizes similarities and connections between lane features, such as origin and destination, which lets it make reliable predictions even when historical data for a particular lane isn’t available. The Market Rate widget in the Rate Intelligence user interface provides numbers for similar historical lanes from the Triumph Network, so you can compare the predicted lane to similar ones. This is especially useful when making predictions for lanes with little or no historical data.

The Confidence level is a prediction of accuracy.

It’s important to understand the capabilities of machine learning. The goal of a machine learning model is not to produce exact predictions 100% of the time, but to minimize the average margin of error across a large data sample.

Rate Intelligence keeps track of the average margin of error (expressed as percentage) in a model’s predictions to ensure that the AI is working as it should.

Since the accuracy of predictions will always show some amount of variation, we use the overall margin of error to assess a model’s performance. Even allowing for these variations, an AI with a low margin of error will still be far and away more accurate than a simple average of rates.

Every Rate Intelligence prediction is accompanied by a Confidence Level.

The Confidence Level tells you how likely a rate is to be accurate. It is determined by the amount and quality of information available to the AI. Confidence Levels usually fall between 50 and 100.

The Confidence Level tells you how much confidence you can safely have in a rate prediction. is intended to take the guesswork out of how hard it will be to cover a load at the given price based on the current market conditions. If the model has enough consistent data to narrow down the spread of potential rates, the Confidence Level will be high. If the model doesn't have the data it needs for an accurate prediction, we recommend being more cautious with quotes.

The Confidence Level is primarily determined by two things:

For every prediction, Rate Intelligence provides a Market Rate prediction and a Verified Buy Rate prediction. The Market Rate prediction is produced by the network model which uses our entire network's load data. The Verified Buy Rate is generated by a model trained to predict a specific brokerage’s rates based on their individual buying behavior and buying power, but using both the brokerage’s data and data the Triumph Network. Since the Market Rate prediction model works with a different data set, Confidence Levels will vary between these two predictions. This lets you compare your rates to those in the Triumph Network.

The Market Rate Confidence Level is not specific to any one brokerage, but represents the likelihood that any brokerage within the Triumph Network will be able to buy a specific lane at a specific price. If a brokerage has a smaller spread of truck rates on a given lane, with more volume consistency, the Verified Buy Rate may have a higher Confidence Level. If a brokerage has no historical volume on a lane (and the model was unable to rely on history from other brokerages that were deemed to be similar or other lanes that would be strongly correlated to the requested lane), where the network has a large quantity of historical volume, the Market Rate Confidence will usually be higher.

Triumph Intelligence recommends taking both network and Verified Buy Rate Confidence into consideration when deciding how to cover a specific lane. We provide a Better Rate icon in our UI to indicate which rate has a higher Confidence Level.

Machine learning lets us predict rates for previously unseen lanes by isolating and filtering for specific load features. The models recognize similarities between certain features in the context of historical rates, which allows them to make reliable predictions even if the lane is new to the brokerage. This is another example of the efficiency of machine learning – a human could easily look up a lane with a similar origin and destination, but a machine learning model can find many kinds of similarities and connections across your entire shipping history very quickly and come up with a faster and more reliable prediction, backed up by more extensive data.

It may seem odd for a brokerage to receive high Confidence on a lane they’ve never moved, but the multitude of data points machine learning models take into account lets them provide an accurate rate prediction with high Confidence.

The Rate Accuracy Report breaks out the margin of error numbers for new lanes, so you can see exactly how well the model is doing at predicting rates for lanes with no historical data.