docs/Appendix.md

A. Portfolio-Level Evidence

All models were evaluated SKU-wise using the bias-aware scoring function:

Score = MAE + |Bias|

This penalizes models that appear accurate but drift directionally— a critical failure mode in fresh categories where bias inflates waste or drives stockouts.

Observed portfolio stability patterns (↓ = more stable)

Tier A — Lower-Noise Forecast Models

Model Family	Mean Stability Score (↓ better)
DynamicOptimizedTheta	66.89
SimpleExponentialSmoothingOptimized	67.31
Chronos2	67.65
Theta	67.68
DynamicTheta	67.69
CrostonOptimized / CrostonClassic	67.88–68.36

Tier B — Acceptable Secondary Models

Model	Score
WindowAverage	68.59
HoltWinters	71.40
Holt	71.84

Tier C — High-Noise / High-Drift Models

Model	Score
SeasonalNaive	76.74
LightGBM	83.91
HistoricAverage	84.07
Naive	88.83
RandomWalkWithDrift	92.74

Interpretation

• Tier-A models produce lower bias and reduced noise at the portfolio level.
• ML (LightGBM), without drivers such as discount, weather, or stockout hours, becomes unstable, overreacting to recent noise.
• Naive and drift models exaggerate noise and create planning churn.

Conclusion: FreshNet dynamics favor noise-dampening methods over signal chasing, particularly when demand structure is heterogeneous.

---

B. SKU-Level Model Decisions

Winner share across all evaluated SKUs:

Tier	Model Families	Share
Tier A	Theta-family, SES/Holt, Chronos2, Croston variants	~65%+
Tier B	WindowAverage, HistoricAverage	~20%
Tier C	LightGBM, Naive, Drift	~15%

Interpretation

• Winners did not cluster around ML models.
• The distribution is skewed toward smoothing-based approaches, particularly in volatile and intermittent SKUs.
• LightGBM wins primarily where behavior is quasi-linear and no external drivers are required.

These patterns reflect model–structure alignment, not algorithmic preference.

---

C. Behavioral Regime Analysis

FreshNet SKUs were segmented into three behavioral regimes. Below are frequently observed stability winners within each regime.

---

1) High-High Regime

(unstable timing + unstable magnitude)

Winning Families
Theta-family models
SES/Holt smoothing
Chronos2
Croston variants (for sparse high-volatility SKUs)

Observed behavior

• These models dampen volatility without flattening structure.
• They avoid overreacting after spikes.
• Chronos2 handles mixed signal patterns without strong oscillation.

LightGBM frequently overfit recent bursts, leading to poor forward stability.

---

2) Low-High Regime

(regular recurrence, unstable amplitude)

Winning Families
Holt-Winters
Theta
Chronos2
Croston variants

Observed behavior

• Seasonal regularity supports Holt-Winters performance.
• Amplitude spikes are absorbed more effectively by smoothing models than ML.
• Chronos2 adapts without repeatedly resetting level after shocks.

---

3) Low-Low Regime

(stable, low-variance items)

Winning Families
SES/Holt/Theta
Historic Average (some SKUs)
Croston (intermittent)

Observed behavior

• Model choice has lower impact in this regime.
• Smoothing models converge to similar baselines.
• Chronos2 is neutral — neither dominant nor harmful.

---

D. Example SKU-Level Decisions (Traceable)

SKU Identifier	Stable Winner
CID0_SID0_PID104…	DynamicOptimizedTheta
CID0_SID0_PID118…	Chronos2
CID0_SID0_PID127…	SES/Holt
CID0_SID0_PID319…	CrostonSBA
CID0_SID0_PID229…	Holt-Winters

Purpose:

• guarantees reproducibility
• shows evidence of regime-matched decisions
• prevents subjective reinterpretation

---

What the Evidence Resolves

---

Technically

The evidence demonstrates that:

• Theta/SES models reduce directional drift, a critical failure mode.
• Chronos2 accommodates mixed structure without aggressive overreaction.
• Croston preserves stability for zero-heavy SKUs.
• LightGBM is unsuitable for fresh categories without driver data.

Stability, when matched to structure, dominates complexity

---

Operationally

A stable, structure-aligned anchor model reduces:

• excessive overrides
• store–planner misalignment
• week-to-week forecast resets
• spiraling exception handling

And enables:

• consistent ordering
• predictable labor and waste planning
• cleaner exception signals

---

Economically

Structure-aligned stability reduces:

• re-forecasting cycles
• waste from positive bias
• stockouts from negative bias
• planning churn and meeting load

These are material cost centers in fresh operations.

---

Deployment Decision

Use Theta-family smoothing and SES/Holt as the default signal where structure is stable. Use Croston methods for intermittent SKUs. Use Chronos2 when demand structure is mixed or uncertain. Introduce LightGBM only once driver data (discounts, stockout hours, weather) is integrated.

Fallbacks are allowed only when:

1. a SKU is structurally deterministic (e.g., controlled replenishment)
2. the category is end-of-life
3. required signals are missing
4. governance mandates a deterministic forecast

All fallback choices must be recorded in the model selection ledger.

---

Closing Position

This evidence shows consistent, structure-conditional patterns, not a single universally dominant model.

Theta/SES, Croston, and Chronos2 remain operationally stable across FreshNet’s volatile, mixed-pattern, and intermittent regimes when applied appropriately.

They produce forecasts that are not only accurate, but steady enough to support durable planning decisions.

That is why they form the anchor set for FreshNet forecasting, under a regime-aware deployment standard.

---