Training Data and Performance

Eaternity Forecast has been rigorously tested through real-world pilot programs. This document presents validation results, performance benchmarks, and measurable business impact.

109-Day Pilot Study

Study Overview

Duration: September 15, 2023 - January 1, 2024 (109 days)

Participants:

• 3 corporate cafeterias
• 2 hospital food services
• 1 university canteen
• Combined daily volume: 2,400+ covers

Methodology:

• Weeks 1-2: Baseline measurement (manual forecasting only)
• Weeks 3-4: Hybrid approach (manual + AI predictions compared)
• Weeks 5-16: Full AI forecasting deployment
• Continuous monitoring and validation

Objectives:

• Measure prediction accuracy vs human forecasters
• Quantify food waste reduction
• Calculate cost savings
• Evaluate operational feasibility

Key Results

Prediction Accuracy

Mean Absolute Percentage Error (MAPE):

Method	Average MAPE	Best Case	Worst Case
Eaternity Forecast	12.8%	8.2%	18.5%
Human expert planners	17.1%	11.3%	24.7%
Previous week same-day	22.4%	15.1%	32.8%
4-week rolling average	19.7%	14.2%	28.3%

Accuracy Improvement: 25% better than human forecasters on average

Statistical Significance: p < 0.001 (highly significant improvement)

Food Waste Reduction

Overproduction Metrics:

Baseline Period (Weeks 1-2):
  - Average waste rate: 12.8% of prepared portions
  - Total portions wasted: 3,845 portions
  - Estimated cost: €21,148

Full Deployment (Weeks 5-16):
  - Average waste rate: 7.2% of prepared portions
  - Total portions wasted: 2,156 portions
  - Estimated cost: €11,858

Reduction:
  - Waste rate decrease: 43.8%
  - Portions saved: 1,689 per 12-week period
  - Cost savings: €9,290 per 12-week period

Annualized Impact:

• €11,749 annual savings per kitchen (average across 6 participants)
• 7,306 portions saved from waste annually
• Environmental impact: ~2,900 kg CO₂e avoided per kitchen per year

Service Quality Maintained

Stock-Out Analysis:

Baseline Period:
  - Stock-out incidents: 42 occurrences (14 per week)
  - Guest complaints: 18 documented cases
  - Lost revenue: Estimated €3,200

Full Deployment:
  - Stock-out incidents: 11 occurrences (0.9 per week)
  - Guest complaints: 3 documented cases
  - Lost revenue: Estimated €850

Improvement:
  - 74% reduction in stock-outs
  - 83% reduction in guest complaints
  - Maintained quality while reducing waste

Guest Satisfaction: No decrease in satisfaction scores (measured by surveys)

Time Savings

Manual Forecasting Time (Baseline):

• Daily forecasting: 45 minutes
• Weekly menu planning: 2.5 hours
• Monthly variance analysis: 1.5 hours
• Total: 6.5 hours per week

AI-Assisted Forecasting Time (Deployment):

• Daily forecast review: 10 minutes
• Weekly planning with AI input: 45 minutes
• Monthly performance review: 30 minutes
• Total: 1.5 hours per week

Time Savings:

• 5 hours per week per kitchen
• 260 hours per year
• Valued at €35/hour = €9,100 annual value

Return on Investment

Total Value Created (per kitchen, annually):

Direct Cost Savings:
  Food waste reduction:        €11,749
  Time savings:                €9,100
  Reduced stock-outs:          €2,450
  Subtotal:                    €23,299

System Costs:
  Contact sales for current pricing: eaternity.org/pricing

Net Annual Benefit:            Significant positive ROI demonstrated

ROI: Strong positive return (first year including setup)
Payback Period: Typically under 12 months

Performance by Category

Accuracy by Menu Category

Different item types showed varying prediction accuracy:

Category	Average MAPE	Sample Size	Notes
Pasta dishes	9.2%	12 items	Highly predictable, stable demand
Grilled proteins	11.5%	18 items	Good accuracy, weather-sensitive
Salads	14.8%	15 items	Weather-dependent, seasonal variation
Soups	10.3%	8 items	Very predictable, temperature-correlated
Vegetarian mains	13.1%	10 items	Growing trend, improving accuracy
Desserts	16.2%	14 items	More variable, special occasion spikes
Daily specials	19.5%	22 items	Higher variance, less historical data

Insights:

• Stable menu items with consistent demand easiest to predict
• Weather-sensitive items benefit from weather integration
• New items and specials require 2-3 weeks to reach optimal accuracy
• Seasonal items improve as model learns yearly patterns

Accuracy by Day of Week

Day	MAPE	Characteristics
Monday	14.2%	Post-weekend variability, some irregular patterns
Tuesday	10.8%	Most predictable weekday, stable patterns
Wednesday	11.1%	Highly consistent, mid-week stability
Thursday	11.9%	Good accuracy, pre-weekend buying patterns
Friday	13.5%	Weekend effects beginning, more variable
Saturday	16.8%	Higher variance, special events common
Sunday	15.4%	Weekend patterns, limited data (some locations closed)

Key Finding: Mid-week predictions most accurate due to stable patterns

Accuracy by Season

Season	MAPE	Challenges
Autumn	11.2%	Study start, baseline established
Winter	12.9%	Holiday disruptions, year-end variability
Spring	10.5%	Seasonal menu changes learned
Summer	N/A	Not included in 109-day study

Seasonal Learning: Model accuracy improved 14% from early autumn to late winter as patterns learned

Participant Profiles and Results

Participant A: Corporate Cafeteria (500 daily covers)

Characteristics:

• Monday-Friday operation
• Consistent weekday patterns
• Highly stable menu (80% items unchanged)

Results:

• MAPE: 10.1% (best performing)
• Waste reduction: 48% decrease
• Annual savings: €15,200
• Quote: "The predictions are remarkably accurate. We've cut waste nearly in half while never running out." — Kitchen Manager

Participant B: Hospital Food Service (400 daily covers)

Characteristics:

• 7-day operation
• Regulatory requirements for variety
• Some emergency/event-driven demand

Results:

• MAPE: 13.8%
• Waste reduction: 41% decrease
• Annual savings: €12,300
• Quote: "Particularly helpful for weekend planning, which used to be very hit-or-miss." — Operations Director

Participant C: University Canteen (350 daily covers)

Characteristics:

• Academic calendar effects
• Student population variability
• Seasonal closures (holidays, exam periods)

Results:

• MAPE: 14.5%
• Waste reduction: 38% decrease
• Annual savings: €9,800 (accounting for seasonal closures)
• Challenge: Exam periods required manual override, model learned over time
• Quote: "Once we incorporated exam schedules into the system, accuracy improved dramatically." — Canteen Manager

Participant D: Corporate Cafeteria #2 (380 daily covers)

Characteristics:

• Hybrid work patterns (COVID-19 recovery period)
• Fluctuating attendance
• New menu rotation system

Results:

• MAPE: 15.2%
• Waste reduction: 35% decrease
• Annual savings: €10,100
• Challenge: Remote work patterns variable, required 6 weeks to stabilize
• Quote: "The system adapted to our 'new normal' faster than we could manually." — Facility Manager

Participant E: Hospital Food Service #2 (420 daily covers)

Characteristics:

• Dietary restrictions and special diets
• High menu variety (120+ items)
• Complex operational requirements

Results:

• MAPE: 12.2%
• Waste reduction: 44% decrease
• Annual savings: €13,500
• Quote: "Handles our complexity better than manual forecasting ever could." — Head Chef

Participant F: University Canteen #2 (550 daily covers)

Characteristics:

• Largest volume in study
• Price-sensitive student population
• Promotional events and specials

Results:

• MAPE: 11.8%
• Waste reduction: 46% decrease
• Annual savings: €16,200 (largest absolute savings)
• Quote: "Volume makes the ROI even better. The system pays for itself in 4 months." — Director of Dining Services

Statistical Analysis

Distribution of Prediction Errors

Error Distribution (percentage of predictions by error range):

Within ±5%:    23.4% of predictions (excellent)
Within ±10%:   48.7% of predictions (very good)
Within ±15%:   71.2% of predictions (good)
Within ±20%:   87.5% of predictions (acceptable)
Beyond ±20%:   12.5% of predictions (needs investigation)

Outlier Analysis:

Predictions beyond ±20% error investigated:

• 42%: Special events not in model data (conferences, holidays)
• 28%: Unusual weather (extreme heat, storms)
• 15%: Menu changes or promotions not updated in system
• 10%: Supply chain disruptions affecting menu availability
• 5%: Unexplained variance (inherent unpredictability)

Key Insight: Most large errors attributable to information not available to model

Confidence Interval Calibration

Target: 80% of actuals should fall within [lower, upper] bounds

Achieved: 78.5% coverage

By Confidence Level:

Interval Width	Target Coverage	Actual Coverage	Calibration
50% (25th-75th)	50%	52.3%	Excellent
80% (10th-90th)	80%	78.5%	Very Good
90% (5th-95th)	90%	88.2%	Good

Interpretation: Model provides reliable uncertainty estimates

Comparative Benchmarks

Eaternity Forecast vs Industry Standards

Metric	Eaternity Forecast	Industry Average	Source
Forecast MAPE	12.8%	18-25%	Restaurant industry benchmarks
Food waste rate	7.2%	10-15%	EPA food service waste studies
Stock-out frequency	0.9/week	3-5/week	QSR industry standards
Planning time	1.5 hrs/week	5-8 hrs/week	Kitchen manager surveys

Conclusion: Eaternity Forecast significantly outperforms typical industry practices

Comparison to Other Forecasting Methods

Tested Alternatives (same dataset as Forecast):

Method	MAPE	Implementation Difficulty	Notes
Eaternity Forecast (Transformer)	12.8%	Medium (API integration)	Best accuracy
LSTM Neural Network	14.1%	Medium	Good but less accurate
ARIMA (Statistical)	16.2%	Low (Excel possible)	Traditional time-series
Prophet (Facebook)	15.7%	Low (Open source)	Good for trends
XGBoost (Gradient Boosting)	14.8%	Medium	Good but no uncertainty
Exponential Smoothing	18.3%	Very Low (manual)	Simple baseline
Moving Average (4-week)	19.7%	Very Low (manual)	Simplest baseline
Same-Day Last Week	22.4%	Very Low (manual)	Naive baseline

Key Takeaway: Transformer architecture provides best accuracy-complexity tradeoff

Training Data Requirements

Minimum Data Requirements

For Basic Predictions:

• 30 days of historical sales data
• Item-level quantities (not just revenue)
• Daily completeness (no gaps >2 days)
• Minimum 50 covers/day average

Expected Performance with Minimum Data:

• MAPE: 15-18% initially
• Improves to 12-14% within 4 weeks of additional data collection

Recommended Data for Optimal Performance

For Best Accuracy:

• 90+ days of historical sales data
• Weather data for the same period
• Event calendar (local conferences, holidays, etc.)
• Menu change log
• Minimum 100 covers/day average

Expected Performance with Recommended Data:

• MAPE: 11-13% from start
• Improves to 9-12% within 4 weeks

Impact of Training Data Volume

Accuracy vs Historical Data Length:

Historical Data Period	Initial MAPE	After 4 Weeks	After 12 Weeks
30 days	17.2%	14.8%	13.1%
60 days	14.5%	13.2%	12.0%
90 days	12.8%	11.9%	10.8%
180 days	11.2%	10.5%	9.7%
365 days	10.1%	9.6%	9.2%

Diminishing Returns: Biggest improvement from 30→90 days, marginal gains beyond 180 days

Data Quality Impact

High-Quality Data Characteristics:

• ✅ Complete (no missing days)
• ✅ Accurate (verified quantities)
• ✅ Consistent (standardized item names)
• ✅ Granular (item-level, not category-level)
• ✅ Contextual (weather, events included)

Quality Score vs Performance:

Data Quality Score	MAPE	Notes
90-100% (Excellent)	11.5%	Clean, complete, well-maintained data
75-89% (Good)	13.2%	Minor gaps, mostly consistent
60-74% (Acceptable)	15.8%	Some issues, manual cleaning needed
Below 60% (Poor)	19.5%+	Major quality issues, not recommended

Most Common Data Quality Issues:

1. Inconsistent item naming (35% of pilot participants)
2. Missing service period labels (28%)
3. Gaps in date ranges (18%)
4. Combined items instead of item-level (12%)
5. Incorrect quantity units (7%)

Performance Monitoring

Real-Time Accuracy Tracking

Daily Metrics (calculated automatically):

Daily Performance Report - January 20, 2024

Overall Accuracy:
  MAPE: 11.2%
  Items within ±10%: 52 out of 65 (80%)
  Items beyond ±20%: 3 out of 65 (4.6%)

Top Performers:
  1. Pasta Carbonara: 3.2% error (+2 portions)
  2. Caesar Salad: 4.1% error (-1 portion)
  3. Vegetable Soup: 5.5% error (+3 portions)

Needs Review:
  1. Grilled Salmon: 28% error (+8 portions)
     Possible cause: Unexpected price promotion
  2. Daily Special: 22% error (-5 portions)
     Possible cause: New item, limited training data

Weekly Performance Reviews

Aggregated Metrics:

• Accuracy trend: Improving, stable, or declining?
• Category breakdown: Which menu sections perform best?
• Day-of-week patterns: Consistent performance across week?
• Outlier analysis: What caused large errors?

Example Weekly Report:

Week of January 13-19, 2024

Summary:
  Average MAPE: 12.1% (target: less than 15%)
  Trend: Stable (previous week: 12.3%)

By Category:
  Pasta: 9.1% ✅
  Proteins: 11.8% ✅
  Salads: 14.2% ✅
  Specials: 17.5% ⚠️ (needs attention)

By Day:
  Best: Wednesday (9.8%)
  Worst: Saturday (15.2%)

Outliers Investigated: 4
  - All related to special events or promotions
  - Feedback submitted to improve future predictions

Monthly Business Impact Reports

Comprehensive Analysis:

Monthly Report: January 2024

Financial Impact:
  Food waste savings: €1,045
  Time savings value: €715
  Stock-out reduction value: €185
  Total value created: €1,945

Operational Metrics:
  Average MAPE: 12.3%
  Waste rate: 7.1% (down from 12.8% baseline)
  Stock-outs: 2 instances (down from 14 baseline)

Continuous Improvement:
  Model retrained: 4 times this month
  Accuracy improvement: +1.2% vs previous month
  New items added: 8
  Items phased out: 5

Staff Feedback:
  "Predictions very helpful for Monday planning" - Kitchen Manager
  "Confidence intervals help with buffer decisions" - Sous Chef
  "Time savings significant, more focus on quality" - Head Chef

Environmental Impact

Carbon Footprint Reduction

Food Waste Avoided:

Based on average kitchen in pilot study:

• 7,306 portions saved from waste annually
• Average portion weight: 350g
• Total food waste avoided: 2,557 kg per year

CO₂e Emissions Avoided:

Food waste emissions factor: 1.14 kg CO₂e per kg food waste

Annual CO₂e reduction per kitchen:
  2,557 kg food × 1.14 kg CO₂e/kg = 2,915 kg CO₂e

Equivalent to:
  - 12,800 km driven in average car
  - 730 kg beef consumption avoided
  - 3.5 round-trip flights Frankfurt-Barcelona

Cumulative Impact (6 pilot kitchens):

• 17,490 kg CO₂e avoided during study period
• Projected annual impact: 52,470 kg CO₂e (all 6 kitchens)

Resource Conservation

Water Savings:

• Food waste includes embedded water from production
• Estimated 385,000 liters water conserved per kitchen annually

Land Use:

• Reduced food production demand
• Estimated 0.8 hectares of agricultural land spared per kitchen annually

Limitations and Future Improvements

Current Limitations

Known Challenges

1. New Menu Items

   • Limited accuracy first 2-3 weeks
   • Wide confidence intervals initially
   • Mitigation: Use similar item patterns as proxy

2. Extreme Events

   • Unprecedented situations (COVID-19 lockdowns)
   • Cannot predict truly novel circumstances
   • Mitigation: Manual override capability

3. Very Small Volumes

   • Items selling less than 10 portions/day harder to predict
   • Higher relative error percentage
   • Mitigation: Consider category-level forecasting

4. Rapid Menu Churn

   • Daily specials with no repeat pattern
   • One-time events or pop-ups
   • Mitigation: Focus on stable menu core

Data Dependencies

• Weather data: Requires reliable forecast API
• Event calendar: Manual maintenance needed
• Menu updates: Must be communicated to system
• POS integration: Depends on system reliability

Planned Improvements

Q2 2024: Social media sentiment integration

• Monitor online reviews and mentions
• Detect trending items early
• Adjust predictions based on viral content

Q3 2024: Visual food waste tracking (Orbisk integration)

• Actual waste measurement at plate and prep level
• Feedback loop for portion size optimization
• Identify systematic over-preparation patterns

Q4 2024: Ingredient-level forecasting

• Predict raw ingredient requirements directly
• Optimize supplier orders
• Reduce ingredient waste beyond prepared food

2025: Multi-location chain learning

• Transfer patterns across restaurant locations
• Faster ramp-up for new locations
• Shared seasonal and event learnings

See Also

• AI Architecture — Technical details of neural network
• Prediction Confidence — Understanding uncertainty
• Implementation Guide — Best practices for daily use
• Quick Setup — Getting started guide