In a new report sent to Rigzone this week, Standard Chartered revealed that it is launching a machine learning model for near-term Brent price forecasting.
Dubbed SCORPIO (Standard Chartered Oil Research Price Indicator), Standard Chartered describes the new launch as a proprietary tree-based model designed to generate a forecast for Brent crude spot prices on a one-week timeframe.
“The machine learning model programmatically collects and analyzes a set of available data and features, weighing the information into a meaningful signal,” Standard Chartered said in the report.
“It allows an element of explainability and aids the decoupling of market sentiment from fundamentals,” it added.
The model incorporates features including high-frequency data points, pricing information for crude and refined products, technical indicators, positioning data, global stocks, implied demand, imports and exports, as well as non-oil data such as USD strength, PMIs and other macroeconomic inputs, Standard Chartered outlined in the report.
“The model also allows us to separate the effects of unexpected events and macro market sentiment from oil-market fundamentals in explaining short-term price moves,” the company said in the report.
The latest iteration of SCORPIO shows a “statistically significant” 67.3 percent directional accuracy, Standard Chartered noted in the report.
“The mean absolute error is lower than one standard deviation of the observations over the last 52 weeks, and the error standard deviation is also lower than the observations standard deviation,” the company added.
Constructing the Model
Explaining the construction of the model in the report, Standard Chartered said it first looked at the most basic set of features – “high-frequency data points used by economists to model supply and demand fundamentals, as well as technical market features”.
“We then added further data sets, including non-oil-specific macroeconomic data. We trialed several alternative data sets. Some showed considerable importance, while others showed little or no explanatory or predictive power for oil prices in the short term,” they added.
“One prominent example was news sentiment data, where news content on chosen topics or keywords is collected and processed to determine if overall sentiment is negative, neutral or positive,” they continued.
“Although programmatic parsing of news reports can provide an edge in high-frequency trading (for example, the model could immediately react to news such as a drone attack on a production facility), this provided no additional value to the model on a daily to weekly basis,” they went on to state.
The company said in the report that overall model performance was measured on back-tested predictive performance for one period ahead.
“Performance is measured using both value-based metrics (mean absolute error) and directional metrics (up, close to zero or down), where classification and evaluation metrics were used,” Standard Chartered said in the report.
“We draw error bands around all predictions based on quantile regression methods (with the same features). A detailed feature report can be generated that explains the derivation of the predicted price movements, grouped by thematic categories,” it added.
“Due to data availability, the timeframe for data sources used is from 2018 onwards. These dates also ensure that the pre-Covid, Covid and post-Covid trading environments are captured, in addition to Russia’s invasion of Ukraine in 2022,” it continued.
Standard Chartered said in the report that it will continue to refine the model over time, “adding new features if they are shown to improve performance”.
Because not all known drivers are captured in reliable data pipelines, a machine learning model can be considered a simplified representation of a subset of known drivers, Standard Chartered said in the report.
“However, it is still vulnerable to so-called ‘black swan’ events – unpredictable events that would not be picked up within our set of indicators and could significantly impact short-term price movements,” the company added.
“In the oil space, such events could include rapidly developing severe hurricanes, geopolitical developments or terrorist acts, producer policy decisions, or broader macroeconomic events such as bank collapses,” it continued.
“From early March to mid-April 2023, two significant ‘black swan’ events occurred that SCORPIO was unable to forecast. However, it was able to accurately forecast price direction from other indicators in the weeks following these events,” the company went on to state.
The black swan events highlighted in the report were the collapse of Silicon Valley Bank and the implementation of additional voluntary output cuts by some OPEC+ members earlier this year.
In a separate report sent to Rigzone this week, Standard Chartered revealed that SCORPIO forecasts a week on week price increase of $2.1 per barrel for front-month Brent to settlement on October 2.
“The upwards forecast would have been greater were it not for speculative positioning; the model is interpreting the strength of the Standard Chartered money-manager positioning index as a pivot point indicator,” Standard Chartered noted in that report.
“SCORPIO also sees USD strength as weighing on the move higher, with the index making a jump from the 87th percentile over the last five years to the 91st percentile,” the company added.
In that report, Standard Chartered highlighted that one of the potential uses of SCORPIO is to indicate whether speculative positioning is so overextended as to become a dominant price factor.
“For the next week, SCORPIO does see positioning and USD strength as drags, but not enough to pull prices lower yet,” the company said in the report.
Prior to the release of Standard Chartered’s SCORPIO model, Rigzone asked several market participants if AI can predict the oil price.
The answer to that question was no, according to Alex Stevens, the Manager of Policy and Communications at the Institute for Energy Research (IER), which describes itself as a not for profit organization that conducts intensive research and analysis on the functions, operations, and government regulation of global energy markets.
Answering the same question, Hussein Shel – the Director, Chief Technologist, and Head of Upstream for Energy and Utilities at Amazon Web Services (AWS) – said, “machine learning and artificial intelligence technologies, including generative AI and similar language models, are not specifically designed for predicting financial markets, including oil prices”.
When Al Salazar, the Senior Vice President at Enverus Intelligence Research (EIR) was asked if AI can predict the oil price, the EIR representative told Rigzone that “AI could have some advantages in terms of data and computing power that conventional forecasters don’t” but added that “one thing that AI could struggle with is correctly timing OPEC actions along with geopolitical driven-supply outages”.
To contact the author, email firstname.lastname@example.org