AI Restaurant Demand Forecasting System Development

Developing an AI-based demand forecasting system for restaurants

Restaurant inventory and staff management directly depend on forecasting guest traffic and orders. Excess inventory → write-offs and losses. Staff shortages → long waits → negative reviews. The ML system predicts cover count and order composition with 85-90% accuracy on most days.

Forecasting tasks

Cover count forecast: How many guests will visit the restaurant in each time slot (breakfast/lunch/dinner, or every 30 minutes).

Revenue per cover: The average bill depends on: day of the week, season, special menus, and audience composition.

Dish demand forecast: What dishes and in what quantities will be ordered? The basis for mise en place is the preparation of ingredients.

Staffing requirement: Cover count → the number of waiters, cooks, and hostesses per slot.

Factors Determining Demand

Forecasting model

features = {
    # Лаги
    'covers_lag_7d_same_slot': covers_7d_ago_same_time,
    'covers_lag_14d_same_slot': covers_14d_ago_same_time,

    # Время
    'day_of_week': dow,
    'time_slot': slot_30min,
    'month': month,
    'is_holiday': holiday_flag,
    'days_since_holiday': days_to_nearest_holiday,

    # Резервации (forward-looking)
    'reservations_for_slot': reservations_made_for_this_slot,
    'reservations_trend': reservations_vs_7d_ago,
    'walk_in_forecast': estimated_walk_in,  # covers - reservations

    # Внешние
    'temperature': temperature,
    'rain_probability': precipitation_probability,
    'nearby_events_score': event_impact_score,

    # Меню
    'special_menu_flag': has_special_menu,
    'promotional_campaign': active_promo
}

model = lgb.LGBMRegressor(n_estimators=300, learning_rate=0.05)

Dish Forecast

Hierarchy: Cover Forecast → Category Mix → Dish Demand

Category Mix:

# Исторически: в обед заказывают 60% основных блюд, 20% закусок, 20% десертов
# В ужин: 50%/25%/25%
# В праздник: +10% к десертам, +5% к алкогольным напиткам
def dish_category_mix(meal_type, day_type):
    base_mix = historical_category_mix[(meal_type, day_type)]
    return apply_seasonal_adjustment(base_mix, current_season)

Top-N dishes forecast: For each popular dish (top 40, >80% of revenue):

• Share in its category: XGBoost
• Features: day of the week, season, menu item, price

Long-tail dishes: we do not forecast individually - group forecast by category.

Waste Reduction

Excess supplies → food waste. The ML system minimizes:

Safe Order Quantity:

SOQ = quantile(forecast_distribution, p=90)  # не медиана, а 90-й процентиль
# Лучше слегка переготовить (можно использовать завтра)
# чем 86'd dish (завершилось блюдо в меню)

Shelf life management:

• Ingredients with a short shelf life → in daily specials, promotions
• Pre-emptive 86: with a low forecast and small quantity → remove from the menu in advance

Food waste tracker: IoT scales for garbage → feedback: if we constantly throw away a specific ingredient → reduce the order.

Staffing Optimization

def staff_needed(covers_forecast, slot_minutes=30):
    """
    Covers → официанты через labor productivity norm
    """
    tables_needed = covers_forecast / avg_covers_per_table
    servers_needed = ceil(tables_needed / covers_per_server)

    # Кухня: covers × avg_dish_per_cover / production_capacity_per_cook
    kitchen_needed = ceil(covers_forecast * avg_dishes / cook_hourly_capacity)

    return {
        'servers': servers_needed,
        'kitchen': kitchen_needed,
        'host': 1 if covers_forecast > 20 else 0
    }

Integration with the scheduling system: R&D Rotamaster, BambooHR, or Jowi — APIs for creating shifts based on the staffing forecast.

POS integration

• iiko, r_keeper, Tillypad: Russian POS — API or SQL for historical data and requests
• Square, Toast, Lightspeed: Western POS — REST API
• Reservation systems: Restaurant capabilities of Yandex/2gis, OpenTable, Resy — API for reservations

Pipeline:

• Daily 07:00: Import yesterday's data from POS
• Recalculation of the forecast for 14 days ahead
• Chef's notice: mise en place plan for tomorrow
• Manager's notice: 3-day staffing plan

Metrics:

• Cover count MAPE: < 10% для следующего дня
• Dish demand MAPE: < 15% для топ-10 блюд
• Food waste reduction: 15-30%
• Labor cost savings: 5-10% through schedule optimization

Deadlines: Basic cover forecast + staffing – 4-5 weeks. A full-fledged system with dish-level forecasting, waste tracker, and POS integration – 3-4 months.