Seasonal Trend And Holiday Decomposition with Loess (STAHL): A Real-Time Approach to Analyzing High-Frequency Alternative Data
Research Paper
A real-time approach to analysing high frequency alternative data
Authors
Benoit Bellone (Head of Research, QuantCube Technology)
Sebastien Daniel (Lead Data Scientist, QuantCube Technology)
Vincent Haller (Data Scientist, QuantCube Technology)
¶
Abstract
This research note exhibits a new seasonal adjustment methodology de- veloped at QuantCube Technology: Seasonal Trend And Holiday decom- position with Loess (STAHL). Derived from the STL procedure introduced by R. B. Cleveland et al. (1990), STAHL aims to be applied to alternative data series at multiple frequencies. It is able to proceed to Working Day Adjustment coping with time-varying or periodic calendars such as the Chinese New Year, to handle structural or periodic missing values, and to provide a point-in-time procedure to generate unrevised trend and sea- sonal components. In a first part, we describe the STAHL framework and its different key innovative steps: spectral identification of multiple sea- sonal frequencies, industrial preprocessing including resampling, seasonal adjustment with missing value handling, specific holiday adjustment and point in time seasonal trend extraction. In the second part, we provide multiple empirical illustrations based on high frequency data. We notably focus on the US Weekly Initial claims series during the outstanding peri- ods of Covid outbreaks which raised critical issues for real-time seasonality extraction. As such we discuss the consequences of our point-in-time (ie. no revision) principle compared to the adjusted series produced by the US Department of Labour. Second, we focus on many different high frequency alternative data series to explore how STAHL deals with multiple season- ality and holidays. We notably focus on periodic missing value treatments and measure the Chinese Lunar Calendar impact on human activity cap- tured through daily measures of NO2 Air pollution.
(Version: December 21, 2023)
Introduction
Macroeconomic nowcasting has increasingly involved the use of various types of massive data over the past decade. Alternative data refer to non-traditional sources of information that can provide unique insights. They can be clus- tered into four major categories: text data, geospatial data, geolocation data, and structured data. Text data can be recovered via multiple sources, mainly through the internet, such as social media data, professional blogs, news arti- cles, job ads, web searches, or hotel and restaurant reviews, for instance.
The availability of high-resolution satellite imagery and the development of deep learning models have led to numerous applications allowing the recovery of various geospatial data from earth observation satellite images, atmospheric data, or radar data. Geolocation data can take the form of shipping traffic, flights, mobility, or vehicle transit numbers. Structured data can encompass prices of goods and services, real estate prices, internet queries, and web traffic.
QuantCube Technology aims to provide a competitive edge by uncovering hidden patterns, detecting emerging trends, and enhancing predictive models by leveraging all those alternative data. Most of these data are produced by QuantCube Technology at a daily frequency, seven days a week, without ex- ception. These time series exhibit a rather short history, oscillating between five and just over ten years. As such, their statistical analysis must address…………………….
Read the research paper in full by downloading a PDF version using the button below
