ADITI GAWLI

VAISHNAVI GOTUKULE
NIKITA MANE
VEDIKA PINGAT
VAISHNAVI VAITY
The Art of Fashion
Forecasting
Fashion forecasting is the process of predicting and analyzing
emerging trends in the fashion industry. It involves
understanding consumer behavior, market dynamics, and the
ever-evolving landscape of style and design. Fashion
forecasting predicts the moods of society and consumers,
along with their behavior and buying habits and bases what
they may release in the coming future off of the forecast.
The Evolution of Fashion
Forecasting
• The evolution of fashion forecasting using machine learning techniques can
be traced back to the introduction of recommender systems in the 1990s.
These systems were based on substantial research in collective intelligence
and provided consumers with recommendations based on grading
frameworks. The first consumer recommendation system, Tapestry, was
created in 1992, which allowed users to rate messages as good or terrible.
• In the early 2000s, e-commerce retailers began employing fashion suggestion
systems, which were primarily in development until 2007-2008. These
systems recommended fashion products to users based on their previous
purchases, much like gadgets and books. With the continuing development of
computer vision algorithms, tailored suggestions that take into account
personal attributes and user evaluations are gaining popularity nowadays.
• Present Day: Advanced technologies, including AI and machine learning, have
further transformed forecasting. Predictive analytics, big data, and trend
algorithms help forecast trends with high accuracy. Additionally, sustainability
and ethical fashion have become significant factors, influencing future trend
predictions.
METHODOLOGY
This type of thing is called reverse image search, so
suppose we have seen a product and we don't know the name
of it so we can use this to find out the similar things that we
have seen. So, when we upload any image on user interface,
our deep learning model will go to the database and find
similar images, and our database is named "image database,"
which contains nearly 45,000 images related to fashion like Tshirts, shoes,
and many more. Fig. 2 depicts the approach
employed in this study.
Working of CNN Model Fashion
Forecasting
Here, we will import a CNN model called ResENT that is
unique because it was trained by some of the best data
scientists and was previously trained on the most well-
known dataset, known as imageNET. It is also a very high
performing CNN model, meaning its accuracy is very good.
If we don’t want to use this model, we have a second
option where we have to train our own model, but doing
so has primarily two disadvantages i.e., it will take a long
time to train a model, and the second is we will never
achieve the accuracy of ResNET, so in the above shot, we
must import a CNN model named ResNET, which is built-in
under Kera's module.
Machine Learning in Fashion Forecasting
Predictive Power Automated Insights Continuous Learning

Machine learning algorithms can analyze By automating the forecasting process, Machine learning models can continuously
analyze vast amounts of data, identify process, machine learning frees up human continuously adapt and improve their
identify patterns, and make accurate human experts to focus on creative and their accuracy as new data becomes
predictions about future fashion trends. and strategic decision-making. available, ensuring relevance and
trends. adaptability.
Data Collection and
Preprocessing
1 Diverse Data Sources 2 Data Cleaning and
Gather data from social media,
Normalization
e-commerce, trend reports, Ensure data quality by
and historical sales patterns to addressing missing values,
gain a comprehensive outliers, and inconsistencies,
understanding of the fashion enabling accurate and reliable
landscape. fashion forecasting.

3 Feature Engineering
Extract relevant features from the data, such as color, silhouette, and
material, to build predictive models that capture the nuances of fashion.
Supervised Learning Techniques
Regression Time Series Analysis
Predict continuous fashion metrics, such as sales volume or trend Forecast future fashion trends by analyzing historical patterns and
popularity, using regression models. seasonality using time series models.

1 2 3

Classification
Classify fashion items or trends into categories, such as high-end,
streetwear, or retro, using classification algorithms.
Architecture Diagram of Fashion Forecasting
AI and Machine Learning Integration Framework
Forecasting Visual Style in Fashion

Fig. 2 shows the details of the network We evaluate our approach on three datasets
architecture for our attribute prediction model. collected from Amazon by [30]. The datasets
The model is composed of 5 convolutional layers represent three garment categories for women
with decreasing filter sizes from 11 × 11 to 3 × 3 (Dresses and Tops&Tees) and men (Shirts). An item
followed by 3 fully connected layers and 2 in these sets is represented with a picture, a short
dropout layers with probability of 0.5. textual description, and a set of tags (see Fig. 3).
Additionally, each convolutional layer and the Additionally, it contains the dates each time the
first two fully connected layers in our model are item was purchased. These datasets are a good
followed by a batch normalization layer and a testbed for our model since they capture real-
rectified linear unit (ReLU). The last attribute world customers’ preferences in fashion and they
Table 1 summarizes the dataset sizes prediction layer is followed by a sigmoid span a fairly long period of time. For all
activation function. We use the cross entropy loss experiments, we consider the data in the time
to train the network for binary attribute range from January 2008 to December 2013. We
prediction. use the data from the years 2008 to 2011 for
training, 2012 for validation, and 2013 for testing
Figure 4: The discovered visual styles on (a) Dresses, (b) Tops &
Tees and (c) Shirts datasets. Our model captures the fine-grained
differences among the styles within each genre and provides a
semantic description of the style signature based on visual
attributes
 Table 3 evaluates the impact of different data
Table 2 compares the forecasting performance of Figure 5 shows the forecasted popularity of four different representations (visual, textual, and combined) on
different models on three datasets: Dresses, Tops & styles from the Tops & Tees dataset using the proposed forecasting performance across three datasets (Dresses, Tops
Tees, and Shirts. The table shows two error metrics, model. Each graph compares the actual popularity (in blue) & Tees, and Shirts).
MAE (Mean Absolute Error) and MAPE (Mean with the forecasted popularity (in green) from 2008 to - Single Modality Methods: Among methods using only one
Absolute Percentage Error), for each model. 2013. type of data (e.g., visual, text, tags), the proposed approach
- Naïve models (mean, last, drift) have relatively - Graphs (a) and (b): The model accurately captures the (using predicted visual attributes) performs the best. The
higher errors. trend and predicts the popularity in 2013, closely matching CNN-based ClothingNet baselines are less reliable because
- Autoregression models (AR, AR+S, VAR, ARIMA) the actual data. they are sensitive to variations like pose and item placement,
generally perform better than naïve models, but - Graphs (c) and (d): These trajectories are more adding noise to the style trajectories.
ARIMA outperforms others within this group. challenging as they show sudden changes in direction - Text vs. Tags: Textual descriptions outperform tags alone,
- Neural network models (TLNN, FFNN) have the around 2012—one goes from a decline to an increase, and likely due to their richer detail in capturing the appearance of
highest errors, indicating they are less effective for the other does the opposite. Despite these abrupt changes, items.
this task. the model successfully predicts the shift, demonstrating its - Combined Approaches: Combining visual attributes with
- The approach labeled "Ours" (the model proposed robustness even with complex patterns. text and tags further improves forecasting accuracy,
by the authors) outperforms all other models across Overall, the figure illustrates that the model can effectively demonstrating that integrating meta-data with visual cues
all datasets, especially in terms of MAPE, making it forecast style popularity, even under conditions where enhances prediction performance.
the most accurate forecasting method in this trends change unexpectedly. Overall, the table emphasizes that leveraging visual
comparison. attributes yields the most reliable fashion forecasts, and
combining them with meta-data strengthens the results
further.
 Figure 7 illustrates the temporal popularity trajectories of six different
styles from the Dresses dataset as predicted by the model. The figure
Fig. 6 shows the style popularity forecasts estimated by baselinesfrom the categorizes the styles into distinct trends:
- Out of Fashion (Fig. 7a): Styles rapidly losing popularity.
three forecasting groups in comparison to our approach. The Naive and NN - Classic(Fig. 7b): Styles that maintain steady popularity with little change
based forecast models seem to over time.
produce larger prediction errors. Our model performs the - Trending (Figs. 7c and 7d): Styles that are gaining popularity quickly.
best followed by the Autoregressor (AR) - Unpopular (Fig. 7e): Styles with low popularity and no signs of
improvement.
- Re-emerging (Fig. 7f): Styles that were once popular, declined, and are
now resurging.
Overall, Figure 7 shows how the model effectively identifies and
differentiates between various style trends, highlighting its ability to track
and forecast fashion dynamics over time.
 Table 4 evaluates the forecasting performance of the model by
varying the number of styles \( K \) between 15 and 85. The key
findings include:

- Effect of Increasing K: As \( K \) increases, more styles are

Fig. 8 shows the forecasted texture and shape attributes introduced, but some do not represent consistent patterns, which
for the Dresses test set. Our model successfully captures adds noise to the time series and raises the forecasting error.
the most dominant attributes in both groups of - Optimal Value of K: At \( K = 30 \), the styles exhibit a coherent
visual appearance with an appropriate level of detail, balancing
attributes, correctly giving the gist of future styles.
between too few and too many styles.
- Error Variance: Despite the increase in error with higher \( K \),
the overall performance variation remains within acceptable
bounds across the tested values.

In essence, Table 4 shows how the choice of \( K \) impacts the

model’s accuracy and highlights the challenge of balancing detail
and consistency in style representation.
Deep Learning in Fashion
Forecasting
Computer Vision Natural Language Processing
Processing
Utilize convolutional neural networks
networks to analyze fashion images Apply deep learning techniques to
images and extract visual features for to process textual data, such as social
features for trend prediction. social media posts and product
descriptions, to understand
consumer sentiment and
preferences.

Generative Models
Generate novel fashion designs and concepts using generative adversarial networks
(GANs) and variational autoencoders (VAEs).
Case Study: Luxury Fashion Brand
Forecasting

Sales Data Runway Trends

Analyze historical sales data to identify best- Monitor runway shows and fashion week
selling products and understand customer events to anticipate emerging trends and
preferences. adapt product offerings.

Social Media Insights Market Research

Leverage social media data to gauge Conduct comprehensive market research to
customer sentiment and spot upcoming research to understand the competitive
style shifts. competitive landscape and identify growth
growth opportunities.
Real-World Applications of
Fashion Forecasting
Product Development
Inform the design and production of new fashion products based on
predicted trends.

Inventory Management
Optimize inventory levels and supply chain operations to meet
anticipated demand.

Marketing Strategies
Develop targeted marketing campaigns that align with forecasted
consumer preferences.
HEURITECH’S FASHION TOOL
Heuritech's Fashion Trend Prediction AI is a sophisticated tool designed to analyze vast amounts of data from various sources to identify
emerging fashion trends.

FEATURE’S: WORKING: APPLICATION’S:

• Data Integration 1. Data Collection • Product Development
• Advanced Analytics 2. Data Processing • Marketing and Advertising
• Trend Prediction 3. Feature Extraction • Inventory Management
• Consumer Insights 4. Model Training • Supply Chain Management
• Competitive Analysis 5. Trend Prediction
The Future of Fashion Forecasting

Generative AI Immersive Experiences Connected Fashion

Advancements in generative AI will enable Integrating augmented and virtual reality The Internet of Things (IoT) will enable real-
the creation of novel fashion designs and reality technologies will provide consumers real-time monitoring of consumer behavior
concepts, revolutionizing the creative consumers with personalized, interactive behavior and preferences, enhancing
process. interactive fashion experiences. fashion forecasting capabilities.
CONCLUSION
The conclusion highlights that the proposed vision-based fashion forecasting
model effectively predicts fashion trends by discovering detailed visual styles
from large-scale data without supervision. The model identifies unique style
patterns, describes them with key visual attributes, and predicts their future
popularity based on user behavior, revealing each style’s life cycle and status
(e.g., trending or declining). The results demonstrate that visual data is
crucial for accurate predictions, outperforming text-based methods. The
model's approach is versatile and can be applied to other domains like
accessories, cars, and furniture, where visual style is relevant.