intent-solutions-io/nixtla-baseline-review

Nixtla Baseline Review Skill

Overview

Analyze baseline forecasting results from the nixtla-baseline-m4 workflow. Interpret metrics, compare models, surface patterns, and recommend next steps.

When to Use This Skill

Activate this skill when the user:

• Asks "Which baseline model performed best?"
• Requests interpretation of sMAPE or MASE metrics
• Wants to compare AutoETS vs AutoTheta vs SeasonalNaive
• Says "Explain these baseline results"
• Needs guidance on model selection based on baseline performance

Prerequisites

• Baseline results must exist in nixtla_baseline_m4/ directory
• At minimum, results_*.csv file must be present
• CSV format: columns series_id, model, sMAPE, MASE

Instructions

Step 1: Locate Results Files

Use the Read tool to find baseline results:

# Check for results directory (use Bash tool)
ls -la nixtla_baseline_m4/

# Identify most recent results file
ls -t nixtla_baseline_m4/results_*.csv | head -1

Expected files:

• results_M4_Daily_h{horizon}.csv - Full metrics table
• summary_M4_Daily_h{horizon}.txt - Text summary (optional)

If files are missing, inform the user they need to run /nixtla-baseline-m4 first.

Step 2: Load and Parse Metrics

Read the metrics CSV file:

# View first few rows to confirm format
head -10 nixtla_baseline_m4/results_M4_Daily_h*.csv

# Or use Read tool to load the full file

Expected CSV structure:

series_id,model,sMAPE,MASE
D1,SeasonalNaive,15.23,1.05
D1,AutoETS,13.45,0.92
D1,AutoTheta,12.34,0.87
D2,SeasonalNaive,18.67,1.23
...

Calculate summary statistics manually or with bash:

• Count total series: cut -d, -f1 results_*.csv | sort -u | wc -l
• Extract model metrics: grep "AutoTheta" results_*.csv

Step 3: Calculate Summary Statistics

For each model (SeasonalNaive, AutoETS, AutoTheta), calculate:

• Mean sMAPE: Average across all series
• Median sMAPE: Middle value (less sensitive to outliers)
• Standard Deviation: Measure of consistency
• Series Won: Count how many series each model performed best on

Step 4: Interpret Metrics

sMAPE (Symmetric Mean Absolute Percentage Error):

• Range: 0% (perfect) to 200% (worst)
• Good: < 10%, Acceptable: 10-20%, Poor: > 20%

MASE (Mean Absolute Scaled Error):

• < 1.0: Better than seasonal naive baseline
• 1.0: Same as seasonal naive

• 1.0: Worse than seasonal naive

• Interpretation: "MASE of 0.85 means this model is 15% better than naive seasonal"

Model Characteristics:

• SeasonalNaive: Simplest baseline, repeats seasonal pattern. Good for stable series.
• AutoETS: Exponential smoothing with automatic parameter selection. Good for trend + seasonality.
• AutoTheta: Theta method with optimization. Often wins on M4 benchmarks, handles trends well.

Step 5: Identify Patterns

Look for:

• Dominant model: Which model wins most often?
• Consistency: Which model has lowest standard deviation in sMAPE?
• Series-specific patterns: Are there series types where certain models excel?
• Failure cases: Where do ALL models struggle? (sMAPE > 30%, MASE > 1.5)

Step 6: Generate Structured Explanation

Provide analysis in this format:

## Baseline Performance Analysis

Based on {N} series from M4-Daily with horizon={H}:

**Overall Winner**: {ModelName}
- Mean sMAPE: {X.XX}% (vs {Y.YY}% for second place)
- Wins on {N} out of {Total} series ({Percentage}%)
- Consistent performance (std dev {X.X}%)

**Key Insights**:
1. {Insight about winner's strengths}
2. {Insight about other models or patterns}
3. {Insight about failure cases or opportunities}

**Recommendation**: Use {ModelName} as production baseline. {Additional context about when to consider alternatives}

**Next Steps**:
- {Actionable recommendation 1}
- {Actionable recommendation 2}

Output

• A structured markdown analysis with a clear winner, key insights, and recommended next steps.

Examples

Example 1: Simple Query

User: "Which model was best?"

Response:

Based on the baseline results (50 series, horizon=14):

**AutoTheta won** with average sMAPE of 12.3%
- Outperformed AutoETS (13.5%) and SeasonalNaive (15.2%)
- Most consistent with std dev of 4.2%
- Recommended for production use

Example 2: Detailed Comparison

User: "Compare AutoETS vs AutoTheta performance"

Response:

## AutoETS vs AutoTheta Comparison

| Metric        | AutoETS | AutoTheta | Winner   |
|---------------|---------|-----------|----------|
| Mean sMAPE    | 13.5%   | 12.3%     | Theta ✓  |
| Median sMAPE  | 12.9%   | 11.8%     | Theta ✓  |
| Std Dev       | 5.1%    | 4.2%      | Theta ✓  |
| Series Won    | 18/50   | 32/50     | Theta ✓  |

**Key Differences**:
- Theta handles trends better (exponential smoothing with drift)
- ETS excels on strong seasonality (state space framework)
- Theta more consistent across diverse patterns

**Recommendation**: Use AutoTheta as default, fallback to AutoETS for highly seasonal data.

Example 3: Metric Interpretation

User: "What does sMAPE of 12.3% mean?"

Response:

sMAPE of 12.3% is **good performance** for forecasting:

- This means on average, predictions are off by about 12% from actual values
- For M4 Daily benchmark, anything under 15% is considered good
- This is in the "acceptable" range (10-20%)

Context: If you're forecasting daily sales of 100 units, a 12.3% sMAPE means your forecast will typically be within ±12 units of the true value.

MASE provides additional context - if MASE < 1.0, this forecast beats a simple "repeat last year" baseline.

Error Handling

If results files are missing:

I don't see baseline results in nixtla_baseline_m4/.

Please run the baseline command first:
Run: `/nixtla-baseline-m4 horizon=14 series_limit=50`

This will generate the metrics files I need to analyze.

If CSV is malformed:

The results file exists but appears malformed. Expected columns:
- series_id, model, sMAPE, MASE

Please re-run /nixtla-baseline-m4 to regenerate clean results.

Resources

For complete technical details, see:

• Architecture: 000-docs/6767-a-OD-ARCH-nixtla-claude-plugin-poc-baseline-lab.md
• Planning: 000-docs/6767-b-PP-PLAN-nixtla-claude-plugin-poc-baseline-lab.md
• Baseline Lab Overview: 000-docs/6767-d-OD-OVRV-nixtla-baseline-lab-overview.md