The Challenge
In the fast-paced world of fashion retail, managing inventory efficiently is crucial. An online fashion retailer was grappling with the challenge of maintaining a balance between supply and demand. Overproduction often led to excessive markdowns, eating into the company's profit margins, while underproduction resulted in missed sales opportunities. The retailer was keen on finding a solution that would help accurately predict demand, thereby reducing unnecessary markdown waste.
The Solution
To tackle this challenge, we at Adyantrix developed a sophisticated machine learning model specifically tailored for the fashion retail industry. Our team of data scientists and retail specialists collaborated closely with the client's internal teams to understand their unique inventory and sales patterns.
The solution was an AI-driven demand forecasting model designed to predict sales with high accuracy. The model utilised historical sales data, seasonality patterns, fashion trends, and external factors such as weather changes to forecast future demand. Advanced machine learning algorithms, including time-series analysis and regression models, were applied to anticipate the optimal inventory levels required to meet customer demand without excessive excess.
The model was seamlessly integrated into the retailer's existing IT infrastructure. We provided training and support to the retailer's staff to ensure they could make the most of the new forecasting capabilities. The solution was further enhanced by real-time data analytics, allowing for quick adjustments to inventory as new data came in.
Key Results
The implementation of the demand forecasting ML model yielded remarkable results. The retailer achieved a 33% reduction in markdown waste, significantly improving their profit margins. With more accurate demand predictions, the retailer was able to optimise their stock levels, reducing the need for drastic markdowns on unsold items.
Furthermore, the retailer experienced a notable increase in overall sales due to improved stock availability. By aligning inventory more closely with customer demand, the retailer ensured that popular items were consistently in stock, thereby capturing sales opportunities they would have previously missed.
The solution not only delivered tangible financial benefits but also enhanced the retailer's operational efficiency. With a more data-driven approach to inventory management, the retailer could focus on strategic growth areas, informed by insights drawn from the forecasting model.
Overall, this project demonstrated the substantial impact of leveraging artificial intelligence and machine learning in the retail industry, specifically within fashion, to drive efficiency and profitability.
Technical Approach
Fashion demand forecasting is a materially harder problem than demand forecasting in most other retail categories because of two compounding factors: extreme seasonality (a winter coat has a selling window of approximately 14 weeks) and the new-product cold-start problem (a new SKU has no sales history at launch). Both required purpose-built modelling strategies.
For the core forecasting engine, we implemented a hierarchical ensemble that combined three model types:
- Facebook Prophet for capturing long-range seasonal patterns and calendar effects (bank holidays, key promotional dates, fashion week periods) at category level. Prophet's additive decomposition made it interpretable enough for the buying team to validate its seasonal assumptions without requiring data science expertise.
- LightGBM gradient boosting at SKU level, incorporating 47 engineered features including sell-through rate at week two (a strong early signal of final sell-through in fashion), colour and size run completeness, Google Trends indices for relevant search terms, and a rolling 12-week competitor markdown signal scraped from five key competitor sites.
- Matrix factorisation for the cold-start problem: New SKUs were embedded in a product attribute space (colour, category, fabric composition, price point) and their initial demand was estimated by proximity to historically similar SKUs in that space. This approach reduced the cold-start error rate by 41% compared to the retailer's existing new-product forecasting method.
The three model outputs were blended using a stacking meta-learner trained on a 26-week held-out validation set, with the blending weights updated quarterly as new validation data accumulated. The entire pipeline — data ingestion, feature engineering, training, validation, and forecast generation — was orchestrated in Apache Airflow running on a managed cloud environment, with model artefacts versioned in MLflow.
Implementation Highlights
The data preparation phase uncovered a significant quality issue: the retailer's historical sales data contained approximately 18 months of records where return transactions had been logged as negative sales rather than as separate return events. This created downward-biased demand signals for several high-return categories, particularly occasionwear. We rebuilt the return-adjusted sales history by reconciling order management and warehouse management system records, which added three weeks to the data preparation phase but was essential for model integrity.
Feature engineering was the most iterative part of the project. The initial feature set of 22 variables produced a mean absolute percentage error (MAPE) of 28% on the validation set — better than the retailer's existing model but not yet meeting our target of sub-20% MAPE. Over six weeks of experimentation, we identified three high-signal features that the initial set had omitted: the number of days remaining in the product's markdown clearance window (a strong predictor of accelerated sell-through), the ratio of the product's current price to the category average price (a relative value signal), and a binary flag for whether the product had been featured in the retailer's weekly email campaign in the prior two weeks (which produced a measurable but short-lived demand spike).
With these additions, the ensemble achieved a 17.3% MAPE on the validation set — comfortably within target and a significant improvement over the retailer's previous 31% MAPE.
Measurable Outcomes
The 33% reduction in markdown waste was the headline result, but the underlying mechanics deserve detail. Before the model, the buying team purchased approximately 12% more stock than anticipated sell-through warranted, as a buffer against forecasting uncertainty. The model's improved accuracy allowed this buffer to be reduced to 6% without increasing stockout events — effectively halving the excess stock exposure. At the retailer's scale, this translated to a reduction in end-of-season markdown inventory value of approximately £4.1 million across the first full trading year.
Stockout rates in the top-selling 20% of SKUs fell by 11 percentage points, from 23% to 12%, because the model's more accurate demand signal allowed the buying team to place more confident initial orders on proven lines rather than hedging with conservative quantities. The consequent improvement in availability on best-sellers contributed an estimated £1.8 million in incremental revenue during the first year.
The model's financial contribution — markdown waste reduction plus incremental revenue — totalled approximately £5.9 million in the first year against a project investment of £420,000, yielding a first-year return of approximately 14x.
Lessons Learned
The most instructive lesson was around model interpretability as a prerequisite for adoption. The initial ensemble produced excellent accuracy metrics but was functionally a black box from the buying team's perspective. Buyers are experienced professionals with well-developed intuitions about their categories; a model they cannot interrogate is a model they will not trust, regardless of its statistical performance. We addressed this by building a companion "what drove this forecast" view in the dashboard that decomposed each SKU's forecast into its top five contributing factors. Once buyers could see that a reduced forecast for a parka was driven by a warm-weather forecast and low week-two sell-through rather than an arbitrary model output, trust and adoption improved significantly.
We also learned that retraining cadence matters as much as model architecture in a fashion context. A model trained in August on summer data will systematically underestimate winter demand if not retrained before autumn buying decisions. We established a mandatory quarterly full retraining cycle and a lightweight incremental update run every two weeks to incorporate the latest sales data, ensuring the model remained calibrated to current trading conditions throughout the year.
Why This Approach Worked
The hierarchical ensemble approach worked because no single model architecture is optimal across the full range of forecasting tasks in fashion retail. Prophet is excellent at long-range seasonality but cannot incorporate SKU-level features; LightGBM can handle rich feature sets but needs sufficient data history to perform well; matrix factorisation handles cold-start scenarios but cannot capture time dynamics. By assigning each model type to the task it handles best and blending their outputs, we achieved accuracy that no single model could have delivered alone.
Equally important was the decision to invest in data quality before model development. It is tempting to begin modelling quickly and treat data quality issues as something to work around. In our experience, the opposite approach — fixing the data first — produces models that are more accurate, more stable, and easier to maintain, because every subsequent retraining run benefits from the clean foundation rather than requiring the same data repair to be repeated.
Speak with our ML Model Development team at Adyantrix to find out how we can support your next project.
Work with Adyantrix
If you are looking to tackle a similar challenge, Adyantrix has the expertise to help across the full project lifecycle. Our ML model development practice covers supervised, unsupervised, and deep learning models. Our AI & machine learning practice covers ML model development, MLOps, and intelligent automation. Our data analytics practice covers BI reporting and self-serve analytics platforms. Our data engineering practice covers pipeline design, streaming, and data infrastructure. Get in touch to discuss your requirements — no commitment required.
