AI Fashion Collection Demand Forecasting

AI-based forecasting of demand for fashion collections

Fashion is one of the most challenging markets to forecast demand for. Short SKU lifecycles (6-12 weeks), high dependence on trends and weather, and a lack of historical data on new SKUs all make traditional planning methods ineffective. Machine learning approaches reduce overstocks and stockouts by 20-35%.

Features of fashion forecasting

Cold Start Problem: New collection – no historical sales. Solutions:

• Attribute-based forecasting: forecasting based on characteristics (color, pattern, category, price segment)
• Transfer learning: similar to last season's article as anchor
• Analogous items: clustering new items to existing SKUs with history

Seasonality + Fashion Trend:

# Decomposition sales signal
# Sales = Seasonal × Category Trend × Fashion Trend × Price Effect × Random
# Fashion Trend: внешние сигналы (Instagram, Vogue, runway)

Short SKU Life: Classic time series require a long history. Instead, we use cross-sectional models at the SKU level.

Data sources

Internal:

• POS data by week: sales, returns, discounts
• Inventory data: balances, out-of-stock dates
• Product characteristics: category, brand, color, material, size, price

External trend signals:

• Google Trends: Search query dynamics by category
• Instagram/Pinterest: Engagement with fashion content (via API or scraping)
• Runway analysis: detecting trends from fashion shows (CV in photos from ModaOperandi, Vogue Runway)
• Weather data: Temperature directly affects jacket/swimsuit sales

Social Listening:

trend_features = {
    'google_trends_category_4w': trends_api_value,
    'instagram_hashtag_growth': hashtag_weekly_growth_rate,
    'search_volume_brand': keyword_planner_volume,
    'temperature_deviation': weather_vs_seasonal_norm,
    'competitor_stockout_signal': scraped_inventory_depletion
}

Forecasting models

Attribute-based LightGBM: For each new product, peak week sales and sell-through rates are predicted based on attributes and trend features. Training is based on historical collections.

Cluster + Analogous Item:

from sklearn.cluster import KMeans

# Кластеризация по attribute embedding
def find_analogous_items(new_item_features, historical_items, n_clusters=50):
    kmeans = KMeans(n_clusters=n_clusters)
    labels = kmeans.fit_predict(historical_items['features'])
    new_cluster = kmeans.predict([new_item_features])[0]
    analogs = historical_items[labels == new_cluster]
    return analogs.sort_values('similarity_score', ascending=False).head(5)

SKU Life Cycle - Sales Curve: Not all SKUs are created equal. Life cycle curve clustering:

• Type A: Fast start → smooth decline (bestseller)
• Type B: slow start → peak at 4 weeks (niche item)
• Type C: smooth sales, basic articles Forecast of curve shape → order distribution over time.

Pre-Season Planning vs. In-Season Adjustment

Pre-Season (6-9 months before the start):

• Initial order based on attribute forecast
• Buy quantities according to the size chart (size curve model)
• Open-to-buy budget by category

In-Season adjustment (weekly): After the first 2-3 weeks of actual sales - Bayesian update of the initial forecast:

def bayesian_forecast_update(prior_forecast, observed_sales, sell_through_weeks):
    """
    Обновление прогноза по первым неделям
    Sell-through rate в первые 2 недели = сильный предиктор финального результата
    """
    early_st_rate = observed_sales / prior_forecast[:sell_through_weeks].sum()
    scaling_factor = early_st_rate ** 0.7  # регрессия к среднему
    return prior_forecast * scaling_factor

Reorder and Markdown triggers:

• If sell-through > 70% in week 4 → reorder (if possible according to production cycle)
• If sell-through < 30% in week 6 → markdown begins according to the Markdown calendar

Size distribution

Size Curve modeling: Historically: XS:S:M:L:XL = 5:20:35:25:15 for this category. ML adjusts for regions, channels, and price segments:

size_curve = lgbm.predict_proba(
    category=category,
    price_tier=price_tier,
    channel=['online', 'store'],
    region=region
)
# → оптимальное соотношение размеров в заказе

Last size problem: Stockout of one size = loss of entire sale. Optimization: small buffer of sizes with the least availability.

Evaluation Metrics

Timeframe: Attribute-based forecasting + analogous item matching + in-season update — 6-8 weeks. A full system with size curve, social trend signals, and Markdown optimization — 3-4 months.