artificial intelligence research and forecasting model

State-of-the-art forecasting: our groundbreaking approach to trend prediction in fashion

Key takeaways

  • Our new forecasting model incorporates the most recent breakthroughs in forecasting technology, featuring state-of-the-art models recently highlighted in scientific research.
  • Leveraging the latest research from Heuritech (David et. al., 2023), our new model refines the use of external fashion signals. This enhancement allows us to more accurately predict and emerging disruptive trends, setting a new standard in fashion forecasting.
  • With its advanced capabilities, we are now setting the stage to develop forecasts that can span 18 to 24 months into the future.

At Heuritech, we quantify and predict consumer demand thanks to the largest dataset on fashion and most accurate forecasting model in the world. Thanks to millions of images analyzed from social media, computer vision and artificial intelligence, we are able to predict fashion trends.


In this article, the Research & Development team at Heuritech (Etienne David, Oscar Bouvier and Paul Morel) wanted to deep dive into our state-of-the-art forecasting model using artificial intelligence. Most importantly, after 2 years of relentless innovation and deep research, we are proud to unveil our new forecasting model.

Up to now, Heuritech proprietary model has harnessed an ensemble of specialized forecasting models.

Our previous  forecasting model included 6 specialized forecasting techniques: 4 statistical models (snaive, stlm, tbats, ets) and 2 cutting-edge deep learning approaches. A complete description of the statistical models are available in Forecasting: Principles and Practice (Hyndman, R.J., & Athanasopoulos, G. 2021). The latter 2 are variants of the model Hermes (David et al., 2022), incorporating fashion external signals to identify disruptive trends.

Despite its strengths, the previous iteration of the Heuritech model faced several limitations:

  • Some specialized models have shown improved results in recent studies, outpacing our older methods.
  • The R&D team highlighted that the Hermes model underutilized the potential of fashion external signals.

Heuritech new forecasting model surpasses previous limitations

For our new forecast model, we leveraged the latest advancements from the forecasting community to create a cutting-edge model. Our improvements over 12-month predictions pave the way for developing solutions that provide even longer forecasting horizons.

Our solution doesn’t simply use one algorithm forecasting model but includes several experts, leveraging each of their strengths. We include the most recent state-of-the-art forecasting models released in the scientific literature.

Our new model is a powerful ensemble of 5 algorithms

Heuritech new model is an ensemble of 5 specialized forecasting models, including 3 of the latest forecasting methods of the scientific literature.

  • Next: in-house HMM-RNN model built by our research team, specialized at detecting disruption using fashion external signals. (Released in september 2023)
  • Patchtst: Best Transformer-based method of the literature. (Released in march 2023)
  • N-BeaTS: Best RNN-based method of the literature. (Released in november 2022)
  • Ets: Standard statistical method commonly used in the literature.
  • Snaive:  Standard statistical method commonly used in the literature.


Depending on the nature of the predicted fashion time series, a weighting will be applied on each 5 models’ forecast to provide the most accurate final predictions.

The key to our success lies in our ability to leverage external fashion signals

Thanks to Heuritech research work (David et. al., 2023), we managed to finer exploit external fashion signals in order to better anticipate disruptive trends.

In addition to including the most recent forecasting models of the literature, recent Heuritech research work (David et. al., 2023) has focused on the inclusion of fashion external signals in our forecasting solution. The objective is to be able to leverage early signals to better detect and anticipate future rising star tendencies.

An outcome of this research work is the development of the model called Next (link). Mixing Deep Learning architectures with Hidden Markov models, this new model has been especially designed to be able to leverage fashion external signals, detect early signals and forecast disruptives tendencies.

Example: Next’s accurate forecast for Retro Football sneakers

A striking example of leveraging external fashion signals is evident in the analysis of Retro Football Sneakers for men in Europe. For this trend, Next’s predictions distinctly stand out, offering a unique perspective compared to other specialized forecasting models.

Next predicts a strong increase in popularity for Retro Football Sneakers, driven by fashion-forward influencers on social media. The sharp increase in activity from these trendsetters at the end of 2022, highlighted in orange on our graph, signals this upcoming explosion in demand. Next successfully analyzes these early indicators to forecast a trend burst in 2023.

When we match Next’s forecasts against the actual 2023 data, we see that Next accurately predicted the rise of Retro Football Sneakers a full year ahead. It was the sole model to foresee this trend shift, adeptly interpreting the early influencer signals that others missed.

Heuritech forecast prediction for EU Male Retro football sneakers

But how do we measure our forecasting model performance?

The Research & Development team at Heuritech used the following methodology to compare the previous forecasting solution with the new one:

  1. Retrieve time series historical signals from Heuritech solution.
  2. Hide the last year of historical data (from 2023-01-16 to 2024-01-08).
  3. Compute for each time series, a prediction on the hidden year.
  4. Compare forecasts with what really happened in 2023 by computing error metrics.
  5. Compute the average error over all time series.
  6. Compare results provided by the different forecasting methods evaluated.

To assess forecast accuracy, we use 3 metrics:

  • Mean Absolute Percentage Error (MAPE): for each week, measure the error between the prediction and the historical signals. Take the percentage of the ground truth and average this error over all weeks forecasted.
  • Mean Absolute Scaled Error (MASE): We measure the forecast absolute mean error and divide it by the absolute mean error of a simple Naive forecasting method.
  • Symmetric Mean Absolute Percentage Error (SMAPE): Measure the absolute error of the forecast, and divide it by the mean of the ground truth and the forecast. Average this over all weeks.

Heuritech Forecasting model results: our new model outperforms industry leaders

In our fashion use case, we rigorously assessed a wide array of forecasting models released in the literature. Among all the tested methods, the Heuritech forecasting model emerged as the clear leader, outperforming all other contenders. 

The final Heuritech forecasting model is the combination of “Next” model and other best-in-class algorithms. The models marked with * have been included in the Heuritech solution.

The closer to zero, the better the model’s performance in predicting actual outcomes accurately.

Forecasting models literature


Overall, our tests proved our enhanced accuracy: the Mean Absolute Percentage Error has been reduced by 5% at 12 months.

To conclude, our latest model integrates state-of-the-art forecasting methodologies, blending deep learning and statistical approaches for comprehensive trend analysis.

The implementation of this new model was a crucial step toward our goal of extending the prediction period. With its advanced capabilities, we are now setting the stage to develop forecasts that can span 18 to 24 months into the future. 

Would you like to learn more about Heuritech forecasting model?

Léa Gossein

About the writer: Léa Gossein, Head of Marketing

Passionate about transformative digital experiences, Léa navigates Heuritech's online landscape, ensuring a seamless blend of AI expertise and digital presence.

Questions or feedback? Email us at info@heuritech.com
heuritech_logo

Get contacted by one of our experts

AI-powered insights to back fashion & sportswear players intuitions

PRESS
Press inquiries: press@heuritech.com

Newsletter

© 2024 ALL RIGHTS RESERVED - HEURITECH | PRIVACY POLICY | TERMS & CONDITIONS | 6 rue de Braque, 75003 Paris