Restaurant Implementation Guide

This guide helps restaurant teams successfully integrate Eaternity Forecast into their daily operations, maximize food waste reduction, and maintain excellent service quality.

Understanding the Transition

From Manual to AI-Powered Forecasting

Traditional Manual Forecasting:

• Chef's intuition based on experience
• Spreadsheet tracking of historical trends
• Time-consuming analysis each week
• Inconsistent accuracy (varies by person)
• Limited consideration of external factors

AI-Powered Forecasting with Eaternity:

• Neural network analysis of all historical patterns
• Automatic integration of weather, events, seasonality
• Instant daily predictions
• Consistent methodology
• Continuous learning and improvement

Key Mindset Shift: Forecast doesn't replace chef expertise—it augments it. The system handles data analysis so your team can focus on quality, creativity, and guest experience.

Daily Workflow Integration

Morning Planning Routine

Recommended Schedule (15-20 minutes total):

1. Review Yesterday's Performance (5 minutes)

Compare actual sales to predictions:

Example Review:
Item: Pasta Carbonara
  - Predicted: 52 portions
  - Actual: 48 portions
  - Variance: -8% (within acceptable range)
  - Waste: 4 portions (minimal)

Item: Grilled Salmon
  - Predicted: 28 portions
  - Actual: 35 portions
  - Variance: +25% (investigate)
  - Stock-out: 2:30 PM (note for future)

Questions to Ask:

• Were variances within expected confidence intervals?
• Did any unexpected events occur (weather change, group cancellation)?
• Were there service quality issues (stock-outs, excessive waste)?
• What can we learn for similar days ahead?

2. Check Today's Forecast (5 minutes)

Review current day predictions:

• Service readiness: Are prep quantities aligned with forecast?
• Confidence levels: Which items have high vs low confidence?
• Special considerations: Weather changes, events, holidays
• Adjustments needed: Override predictions if you have information the model doesn't

Dashboard Quick View:

TODAY: Wednesday, January 20, 2024 | Temp: 8°C | No special events

High Confidence Items (>90% accuracy):
  - Pasta Carbonara: 52 portions (range: 48-56)
  - Caesar Salad: 31 portions (range: 28-34)

Medium Confidence Items (75-90%):
  - Grilled Salmon: 28 portions (range: 23-33)
  - Vegetable Curry: 41 portions (range: 36-46)

Low Confidence Items (less than 75%):
  - New Item: Mushroom Risotto: 15 portions (range: 8-22)
    Note: Insufficient historical data

3. Plan Tomorrow (5-10 minutes)

Use next-day forecasts for:

Ingredient Ordering:

• Calculate total quantities needed based on predictions
• Add buffer for high-variance items using upper confidence bound
• Reduce buffer for highly predictable items
• Account for prep waste and portion standardization

Staff Scheduling:

• Align kitchen staff with expected volume
• Prepare for predicted busy vs quiet days
• Plan prep tasks during lower-demand periods

Menu Adjustments:

• Consider pausing items with consistently low predictions
• Promote items with high predicted demand
• Prepare for seasonal transitions

Weekly Planning Routine

Monday Morning Review (30-45 minutes):

1. Week-in-Review Analysis

Aggregate last week's performance:

Week of January 13-19, 2024:

Overall Accuracy: 89.2%
  - Improved from previous week: 85.1%

Top Performers:
  - Pasta Carbonara: 95% accuracy
  - Caesar Salad: 93% accuracy

Needs Attention:
  - Friday fish special: 68% accuracy
    Reason: Supplier delivery delays affected availability

2. Upcoming Week Forecast Review

Look ahead at the week:

• Pattern identification: Are there predictable busy/quiet days?
• Special events: Conferences, holidays, weather patterns
• Menu planning: Should you adjust offerings based on predictions?
• Inventory planning: Bulk ordering opportunities for high-demand items

3. Team Communication

Share insights with kitchen staff:

Example Team Brief:

Week Ahead (Jan 20-26):

Expected Busy Days:
  - Friday (Jan 24): Conference at nearby hotel
    Volume prediction: +35% vs normal Friday

Expected Quiet Days:
  - Monday (Jan 20): Post-holiday slowdown
    Volume prediction: -15% vs normal Monday

Menu Notes:
  - Vegetable Curry showing upward trend (winter comfort food)
  - Salads slightly down (cold weather pattern)
  - Consider promoting soup offerings

Using Confidence Intervals

Every prediction includes a confidence interval showing the range of likely outcomes, not just a single number. This helps you make better preparation decisions based on your priorities.

Strategy	When to Use	Prepare To
Conservative	Expensive/perishable items	Lower bound
Balanced	Standard menu items	Point estimate
Aggressive	Signature/low-cost items	Upper bound

Quick Example:

Pasta Carbonara - Saturday:
  Point estimate: 52 portions
  Confidence interval: 48-56 (80% confidence)

  → Conservative: Prepare 48, keep ingredients for more
  → Balanced: Prepare 52 portions
  → Aggressive: Prepare 56 portions

For detailed strategies, buffer calculations, and decision frameworks, see Understanding Confidence Intervals.

Handling Special Situations

New Menu Items

Challenge: No historical data = low confidence predictions

Initial Strategy:

Week 1: Launch with conservative quantity
  - Use similar item predictions as rough guide
  - Prepare to lower bound of estimated range
  - Monitor closely and adjust daily

Week 2-4: Rapid learning phase
  - Model begins learning actual patterns
  - Confidence intervals narrow
  - Predictions improve significantly

Week 5+: Normal forecasting
  - Sufficient data for reliable predictions
  - Standard confidence levels achieved

Example:

New Item: Mushroom Risotto (launched Monday)

Monday (Day 1):
  - No prediction available
  - Prepare: 20 portions (estimated)
  - Actual: 15 portions sold

Tuesday (Day 2):
  - Prediction: 15 portions (range: 10-20)
  - Confidence: Low
  - Prepare: 15 portions

Week 4:
  - Prediction: 22 portions (range: 19-25)
  - Confidence: Medium (80%)
  - Prepare: 22 portions

Seasonal Menu Changes

Challenge: Menu transitions between seasons

Strategy:

Phase-Out Period (2-4 weeks before menu change):

Declining summer items:
  - Model detects downward trend
  - Predictions gradually decrease
  - Confidence intervals may widen (fewer recent sales)
  - Reduce prep quantities accordingly

Phase-In Period (2-4 weeks after menu change):

New autumn items:
  - Start with low-confidence predictions
  - Use previous year's data if available
  - Rapid learning as sales data accumulates
  - Stabilization within 3-4 weeks

Example Seasonal Transition:

Grilled Vegetable Salad (Summer → Autumn):

Late August:
  - Prediction: 45 portions/day (high confidence)

Early September:
  - Prediction: 32 portions/day (model detects decline)
  - Temperature cooling affecting salad demand

Late September:
  - Prediction: 18 portions/day (preparing to phase out)
  - Consider replacing with autumn item

Butternut Squash Soup (New Autumn Item):

Early September:
  - Week 1 prediction: 25 portions (low confidence, range: 15-35)
  - Week 2 prediction: 28 portions (confidence improving)

Late September:
  - Prediction: 32 portions (medium confidence, range: 28-36)
  - Model has learned demand pattern

Special Events

Challenge: Unusual circumstances not in historical data

Types of Events:

Known Events (Model Aware)

Events the model can detect:

• Public holidays: Christmas, Easter, New Year
• Weather patterns: Temperature, precipitation, seasonal trends
• Day of week: Weekends, Mondays, Fridays
• Local events: If included in your data (conferences, festivals)

How to handle:

• Trust predictions—model has learned these patterns
• Review historical accuracy for similar events
• Minor manual adjustments only if you have specific knowledge

Unknown Events (Override Needed)

Events requiring manual adjustment:

• Unexpected group bookings: Large reservations made recently
• Marketing campaigns: New promotions not in historical data
• Supply issues: Menu substitutions due to ingredient availability
• Facility issues: Construction, equipment failure, limited service

Override Process:

Via dashboard or API:

POST /api/forecast/predictions/override
{
  "date": "2024-01-25",
  "item": "Pasta Carbonara",
  "override_quantity": 75,
  "reason": "Conference group booking (50 pax confirmed)",
  "preserve_ratios": true
}

Via Necta interface:

1. Navigate to forecast for specific date
2. Click "Override" next to item
3. Enter adjusted quantity and reason
4. Save—system adjusts related items proportionally if requested

Example:

Normal Saturday prediction: 52 Pasta Carbonara

Situation: Wedding party of 80 confirmed, 60% selecting Pasta
Expected from wedding: 48 portions
Normal Saturday demand: 52 portions
Overlap (wedding guests = regular guests): -15 portions

Adjusted prediction:
  52 (normal) + 48 (wedding) - 15 (overlap) = 85 portions

Override to: 85 portions
Reason: "Wedding party, 80 pax, 60% pasta selection"

Unexpected Variances

When actual sales differ significantly from predictions:

Step 1: Document the Event

Record details immediately:

Date: January 20, 2024
Item: Grilled Salmon
Predicted: 28 portions
Actual: 42 portions
Variance: +50%

Circumstances:
  - Food critic review published this morning
  - Social media mentions increased 300%
  - Walk-in traffic +40% vs normal Wednesday

Step 2: Provide Feedback to Model

Via dashboard feedback form:

Variance Report:
  - Item: Grilled Salmon
  - Date: 2024-01-20
  - Actual vs Predicted: 42 vs 28
  - Root cause: Media coverage/food critic review
  - Recurring?: No, one-time event
  - Action taken: Prepared additional 14 portions mid-service

Step 3: Learn for Future

Model improvements:

• Immediate: Your feedback helps model understand unusual events
• Short-term: Similar patterns detected faster in future
• Long-term: Model learns to identify early signals (social media trends)

Your process improvements:

• Monitor social media mentions
• Track local food press coverage
• Build buffer processes for surge demand
• Maintain relationships with suppliers for emergency orders

Team Training and Adoption

Training Kitchen Staff

Week 1: Introduction

Goals:

• Understand what Forecast is and how it works
• Learn to access predictions
• Identify team roles and responsibilities

Activities:

• 30-minute overview presentation
• Dashboard walkthrough
• Q&A session
• Distribute quick reference guides

Key Messages:

"Forecast is a tool to help us, not replace us"
"We still need your expertise for final decisions"
"The system learns from our feedback"
"Goal is less waste and better service, not perfect predictions"

Week 2-3: Hands-On Practice

Goals:

• Practice using predictions in planning
• Learn confidence interval interpretation
• Develop feedback habits

Activities:

• Daily 5-minute stand-up meetings
• Review previous day's accuracy together
• Discuss variances and learn together
• Encourage questions and experimentation

Practice Scenarios:

Scenario 1: High confidence prediction
  - Item with 95% accuracy, narrow confidence interval
  - How much do we prepare?

Scenario 2: Low confidence prediction
  - New item, wide confidence interval
  - What's our strategy?

Scenario 3: Unexpected event
  - Large group booking after forecast generated
  - How do we adjust?

Week 4+: Continuous Improvement

Goals:

• Optimize workflow integration
• Track performance improvements
• Share successes and challenges

Activities:

• Weekly team review meetings
• Monthly accuracy and waste reduction reports
• Celebrate wins (waste reduction, time saved)
• Iterate on processes

Overcoming Common Resistance

"I don't trust the predictions"

Response:

• Start with observation mode (compare predictions to current method)
• Show accuracy data from validation studies
• Begin with low-stakes items
• Track comparative performance

Action:

Week 1: Run parallel forecasting
  - Your method vs Forecast
  - Track which is more accurate
  - Review results together weekly

Most teams find:
  - AI matches or beats their accuracy within 2 weeks
  - Time savings become immediately apparent
  - Confidence builds with exposure

"The system doesn't understand our kitchen"

Response:

• Explain how model learns YOUR specific patterns
• Show examples of pattern detection (weekday/weekend, seasonality)
• Highlight continuous learning from feedback

Action:

- Review training data together—it's YOUR historical sales
- Show how confidence improves as model learns
- Demonstrate that feedback directly improves future predictions
- Point out kitchen-specific patterns model has learned

"What if the predictions are wrong?"

Response:

• No prediction is perfect (neither are manual forecasts)
• Goal is better average accuracy, not perfection
• Confidence intervals account for uncertainty
• Override capability maintained

Action:

- Calculate current waste from over-forecasting
- Calculate current stock-outs from under-forecasting
- Track Forecast performance vs current method
- Show gradual improvement over time

Measuring Success

Key Performance Indicators

Food Waste Reduction

Metric: Percentage reduction in wasted portions

Tracking:

Before Forecast (Baseline Month):
  - Total portions prepared: 3,450
  - Total portions wasted: 425
  - Waste rate: 12.3%

After Forecast (Month 3):
  - Total portions prepared: 3,280
  - Total portions wasted: 245
  - Waste rate: 7.5%

Improvement:
  - Waste reduction: 39% fewer wasted portions
  - Over-preparation reduction: 5% fewer portions prepared
  - Maintained quality: Zero increase in stock-outs

Cost Savings:

Average portion cost: €5.50
Wasted portions reduced: 180/month
Monthly savings: €990
Annual savings: €11,880

Forecast Accuracy

Metric: Mean Absolute Percentage Error (MAPE)

Calculation:

MAPE = Average of: |Actual - Predicted| / Actual × 100%

Example week:
  - Monday: |48 - 52| / 48 = 8.3%
  - Tuesday: |55 - 51| / 55 = 7.3%
  - Wednesday: |42 - 45| / 42 = 7.1%
  - Thursday: |58 - 54| / 58 = 6.9%
  - Friday: |73 - 71| / 73 = 2.7%

Week average MAPE: 6.5% (excellent)

Target Benchmarks:

• Excellent: less than 10% MAPE
• Good: 10-15% MAPE
• Acceptable: 15-20% MAPE
• Needs improvement: >20% MAPE

Time Savings

Metric: Hours saved per week on forecasting

Typical Savings:

Manual forecasting (before):
  - Weekly menu review: 2 hours
  - Daily forecasting: 30 min × 6 days = 3 hours
  - Variance analysis: 1 hour
  Total: 6 hours/week

With Forecast (after):
  - Automated predictions: 0 hours
  - Daily review: 10 min × 6 days = 1 hour
  - Weekly performance review: 30 minutes
  Total: 1.5 hours/week

Time savings: 4.5 hours/week (75% reduction)
Value: 4.5 hours × €35/hour = €157.50/week
Annual value: €8,190

Service Quality

Metric: Stock-out frequency

Tracking:

Before Forecast:
  - Stock-outs per week: 3-4 instances
  - Guest disappointment: Moderate

After Forecast:
  - Stock-outs per week: 0-1 instances
  - Guest disappointment: Minimal

Improvement: 75% reduction in stock-outs

Monthly Reporting

Executive Summary Template:

Eaternity Forecast Performance Report
Month: January 2024

Key Metrics:
✅ Food Waste: 7.2% (down from 12.8% baseline)
✅ Forecast Accuracy: 91.5% average (target: >85%)
✅ Cost Savings: €1,045 this month
✅ Stock-outs: 2 instances (down from 14 in baseline)

Top Performing Items (>95% accuracy):
  - Pasta Carbonara: 96.2%
  - Caesar Salad: 95.8%
  - House Bread: 95.1%

Items Needing Attention (less than 85% accuracy):
  - Friday Fish Special: 78.3%
    → Action: Review supplier consistency
  - Weekend Brunch Items: 82.1%
    → Action: Longer historical period needed

Total Value Generated: €1,730
  - Waste reduction: €1,045
  - Time savings: €685

Cumulative Savings (3 months): €4,850
Projected Annual Savings: €19,400

Best Practices

Do's

✅ Review predictions daily — Make it part of morning routine ✅ Understand confidence intervals — Use ranges for decision-making ✅ Provide feedback on variances — Help the model learn ✅ Track performance metrics — Measure waste, accuracy, savings ✅ Trust the data — Let go of pure intuition gradually ✅ Communicate with team — Share insights and wins ✅ Maintain override capability — Use when you have special knowledge ✅ Start conservative — Begin with lower bounds until confidence builds

Don'ts

❌ Don't ignore wide confidence intervals — They signal uncertainty ❌ Don't override without documentation — Track reasons for learning ❌ Don't expect perfection — Focus on average improvement ❌ Don't skip training period — Model needs time to learn your patterns ❌ Don't forget to update menu changes — Keep system informed ❌ Don't resist new items — System learns quickly with data ❌ Don't ignore feedback requests — Your input improves accuracy ❌ Don't change planning process too quickly — Gradual transition works best

Troubleshooting Common Issues

Predictions Seem Too High

Possible Causes:

• Recent menu changes not reflected in forecast
• Seasonal decline beginning
• Quality or service issues affecting demand
• Competition or market changes

Actions:

1. Check if item is being phased out seasonally
2. Review recent guest feedback for quality issues
3. Verify menu prices haven't increased
4. Look for new competition nearby
5. Provide feedback to model with context

Predictions Seem Too Low

Possible Causes:

• New promotion or marketing campaign
• Improved quality or recipe changes
• Seasonal uptick beginning
• Positive media coverage

Actions:

1. Verify no recent marketing not in model's data
2. Check for social media mentions or reviews
3. Review if quality improvements made
4. Look for seasonal trend beginning
5. Override temporarily and provide feedback

Accuracy Declining Over Time

Possible Causes:

• Menu changes not updated in system
• Data quality issues
• Significant operational changes
• Model needs retraining

Actions:

1. Verify menu updates are being communicated
2. Check data submission consistency
3. Review for any process changes
4. Contact support for model refresh

Wide Confidence Intervals

Possible Causes:

• High natural variability in demand
• Insufficient historical data
• Multiple unpredictable factors
• New or seasonal items

Actions:

1. Accept higher uncertainty for these items
2. Use conservative approach (prepare to lower bound)
3. Maintain higher ingredient buffer
4. Provide more feedback to accelerate learning

See Also

• Quick Setup Guide — Initial setup and configuration
• Prediction Confidence — Deep dive on confidence intervals
• Features — All available capabilities
• Performance Study — Validation results and benchmarks
• Troubleshooting — Technical issues and solutions