Data Science Driving Efficiency in Wholesale Fuel Pricing

Key Takeaways

• Utilizing historical data, market trends and external factors to precisely forecast fuel demand can assist wholesalers in optimizing inventory management and pricing strategies.
• The integration of data science with market intelligence equips stakeholders with advanced foresight and adaptability, facilitating scalable pricing and revenue optimization in wholesale fuel markets.

The Wholesale Fuel Distribution Ecosystem

Within the expansive network of fuel distribution, terminals stand as essential pillars strategically situated along pipelines, orchestrating the complex interplay of supply and demand. These terminals serve as vital nerve centers responsible for receiving, storing and dispatching fuel to destinations spanning vast geographical areas. At the core of this intricate ecosystem lies the wholesale rack – a nexus where fuel is loaded onto trucks, initiating journeys that fuel our vehicles, power our industries and drive our economies forward.

The term “rack” stems from the loading rack, where trucks directly receive fuel from suppliers, marking the inception of a journey that traverses the intricate maze of wholesale pricing. Transactions at these racks typically involve quantities tailored for truck and trailer transport, with an average load comprising around 8,000 gallons of fuel. However, beneath the surface of these transactions lies a complex framework of interactions, spanning refineries, wholesalers, terminal operators, and retailers or bulk buyers.

Fuel pricing at the wholesale rack is a multifaceted affair influenced by myriad factors, each exerting its own gravitational pull on the final cost. The spectrum of fuel types available – ranging from gasoline and distillate to biodiesel, renewable diesel, ethanol and jet fuel – adds complexity to the pricing calculus. Additionally, costs associated with crude oil, refining processes, transportation logistics, taxes and fluctuating market demand contribute significantly to shaping the pricing landscape. The pricing mechanism often adheres to a formula of “spot + freight + other,” in which the spot price mirrors the commodity price in the refinery’s serving marketplace, freight encapsulates transportation costs and “other” encompasses refining expenses and miscellaneous charges.

To provide a concrete example, let’s consider the city of Atlanta, which is supplied by the Gulf Coast spot market. Assuming a refinery’s cost to refine diesel from the Gulf Coast to Atlanta is approximately $0.035 per gallon, coupled with additional charges amounting to $0.015 per gallon, and considering the spot price for U.S. Gulf Coast Ultra-Low Sulfur No 2 Diesel (I:USGCDSP) at $2.60 per gallon (as of March 25, 2024), the total cost for transporting diesel from the refinery to the Atlanta rack would total $2.65 per gallon ($2.60 + $0.035 + $0.015). Incorporating historical profit margins from refining, averaging about 3-6 cents per gallon, the fuel would likely sell on the market at approximately $2.70 per gallon – a carefully calculated price tag designed to navigate the delicate balance of supply and demand dynamics. Beneath the surface of seemingly straightforward price tags and profit margins lies a domain in which data science holds the promise of enhancing market efficiency and significantly boosting the financial performance of refiners.

Optimizing Wholesale Fuel Pricing Dynamics with Data Science

The mechanics of the rack – akin to a vast reservoir where refiners deposit their fuel and vie to capture market share – present a unique challenge. Despite customers effectively purchasing the same type of fuel from the same storage container, the seller from whom they buy can significantly impact the price. Typically, customers opt for the seller offering the lowest price at the rack. This dynamic creates a scenario in which the seller with the lowest price tends to sell out first, potentially leaving opportunity unrealized for those with higher prices who also manage to sell their inventory on the same day. Conversely, sellers with the highest prices may struggle to make sales. Ideally, a seller aims to set a price that ensures they sell their fuel last, exhausting the inventory they intended to sell, by the end of the day. Given the complex interdependence of numerous variables, conditions and parameters, attaining this delicate equilibrium presents an ideal challenge for data science.

From a data science perspective, addressing the pricing challenge involves dissecting it into several essential components. One critical task is forecasting demand at a specific terminal for various fuel grades, which serves as a foundational aspect of wholesale pricing. This endeavor necessitates a comprehensive analysis of numerous factors including population density, economic indicators, seasonal fluctuations, industrial activity, weather patterns and global events. Fortunately, the data required to support these models is readily accessible and provides a wealth of insights ready to be explored.

Additionally, to fully address the purchasing dynamics, predictions are needed for competitor fuel prices and to estimate the fraction of fuel demand captured based on price rank. Although these represent distinct challenges, they can effectively be viewed as another set of predictive tasks that rely on factors such as spot price, transportation costs and refinery efficiency. Ultimately, this complex problem can be approached as a series of forecasting challenges leveraging readily available data resources that data scientists can meticulously structure, clean and preprocess to ensure the readiness for analysis.

Advanced Data Science Techniques and Solution Framework

Following data preparation, the next crucial step involves feature engineering, in which pertinent features or drivers for price variability are identified and crafted from the collected data. These features, including lagged values, moving averages, seasonal patterns, and extreme and rare events, play a pivotal role in enabling forecasting models to capture underlying trends and relationships in the data. Subsequently, data scientists can carefully select appropriate models tailored for forecasting tasks, which may include time series models, regression techniques, machine learning algorithms or deep learning architectures, aligning with the specific characteristics and nuances of the data.

Amid the array of data science techniques available, the transformer architecture emerges as a standout solution, renowned for its exceptional efficacy in processing sequential data laden with long-range dependencies. Initially developed for natural language processing applications, the transformer architecture has showcased remarkable versatility across diverse domains, including the realm of time series forecasting.

A key advantage of the transformer architecture lies in its adeptness at modeling intricate long-range dependencies within data sequences – an important consideration in predicting fuel prices, which has a rather large number of often highly correlated variables and is particularly significant to extreme weather or geopolitical events. Powered by self-attention mechanisms, this architecture enables the capture of nuanced relationships between distant elements within input sequences, its adaptability to diverse time series patterns and capability to integrate multimodal inputs facilitating the acquisition of complex patterns and dependencies intrinsic to time series data.

Furthermore, the transformer architecture boasts inherent advantages in parallelization and scalability. Unlike traditional recurrent neural networks (RNNs) that also process sequential data, the transformer architecture supports parallel processing of input sequences, leading to expedited training and inference times. This scalability renders it well suited for handling voluminous time series data typical of fuel demand forecasting scenarios.

Integrating the Data Science Outputs

To demonstrate this concept and integrate the three predictions, let’s consider a scenario at a terminal with four sellers offering a specific grade of fuel: assuming that data science methods have accurately predicted competitor prices of $2.68, $2.70 and $2.72 for competitors C1, C2 and C3, respectively, along with estimated price ranks 1, 2, 3 and 4 to sell 60,000, 40,000, 20,000 and 5,000 gallons of fuel, respectively. In this scenario, if a seller has 25,000 gallons of fuel to sell, setting a price just below $2.72 (corresponding to price rank 2) can optimize the selling price while ensuring a complete sellout for the day – a strategic objective established during the problem formulation stage.

Within the wholesale fuel pricing sector, even minor price adjustments of a few cents can trigger significant impacts across the entire supply chain. By combining the capabilities of data science with market intelligence, this fusion becomes a powerful tool that empowers industry stakeholders with invaluable foresight and flexibility essential for navigating a constantly evolving landscape. This enhanced agility allows stakeholders to optimize revenue and profitability on a large scale, adapting swiftly to market dynamics and maximizing outcomes.