Real World Time Series Benchmark Datasets with Temporal Distribution Shifts Global Crude Oil Asset Price and Volatility


Pranay Pasula
UC Berkeley EECS

Abstract

The scarcity of task-labeled time-series benchmarks in the financial domain hinders progress in continual learning and AI algorithm development. Addressing this deficit would foster innovation in this area.

Therefore, we present COB, crude oil benchmark datasets. COB includes 30 years of asset prices that exhibit significant distribution shifts and comes with corresponding task, or regime, labels based on these distribution shifts for the three most important crude oils in the world: West Texas Intermediate (WTI), Brent Blend, and Dubai Crude.

Time-series data pose challenges such as correlatedness, non-stationarity, missing data, and spurious outliers. We address the problem of out-of-distribution detection and emphasize the relevance of our benchmarks for continual learning.

Our contributions include creating real-world benchmark datasets by transforming asset price data into volatility proxies, fitting models using expectation-maximization, generating contextual task labels that align with real-world events, and providing these labels to the public.

We hope these benchmarks accelerate research in handling distribution shifts in real-world data, in particular due to the global importance of the assets represented by the datasets.


Download Benchmark Datasets with Task Labels

Approach

We meticulously transform, fit the data, and transform the results to obtain labels that partition the data into task, or regimes, based on asset price volatility, a macroeconomic trend that correlates with critical global events, such as economic downturns or recessions.

Task label generation on Brent Blend crude oil data. Gray represents recessions per the National Bureau of Economic Research (NBER).

We evaluate four continual learning algorithms: (1) MOLe, (2) MoB, (3) MAML k-shot, and (4) MAML Continuous on all three datasets.

Two of these algorithms, MoB and MOLe, instantiate new few-shot models adapted from a meta-learned prior. The two others, MAML k-shot and MAML Continuous use a single model that is adapted from a meta-learned prior.

Results

1. The benchmark datasets (WTI and Brent resampled weekly, Dubai remained monthly) and task labels can be downloaded here.

2. We report the Percent Improvement in MSE of each algorithm on each benchmark dataset over three Forecast Lengths.

The task labels our algorithm generates improve model accuracy universally over four continual learning algorithms and three forecast lengths

Conclusion

In conclusion, we have introduced three novel time-series benchmark datasets that encompass the global crude oil market and generated task labels aligning with significant real-world events, creating a rich context for further exploration and understanding.

Our goal is to fuel advancements in areas such as continual learning and out-of-distribution detection, while addressing challenges posed by sequential data.

We believe that these benchmarks, due to the societal importance of the assets they represent, can make a substantial contribution to AI research, particularly in handling real-world data that contains distribution shifts.

As AI continues to permeate every sector, the importance of such representative, real-world benchmarks will continue to grow.