I built a forecasting pack for Greek sheep milk production because I was tired of discussing the sector as if the only variables that mattered were “a few cents per kg” and “Sheep Pox Spread.” Price and zoonoses matter, but without volume discipline and without understanding where volume is shifting, price becomes background noise.
This is not an academic exercise. I wanted something I can use operationally: planning, milk-zone governance, logistics, capacity, feta cheese and working capital. The core principle is simple: if I forecast Greece and its regions (prefectures), the system must be coherent- the national forecast must equal the sum of the sub-national forecasts.
Executive snapshot
| Scope | Monthly sheep milk volume (Greece + prefectures), 2020–2025/10; rolling-origin backtest |
| Champion model | ETS (Exponential Smoothing), WAPE ≈ 2.30% |
| Baseline “floor” | Seasonal Naive (SNAIVE), WAPE ≈ 2.59% |
| 2026 corridor | ~734m kg (conservative / floor) to ~768m kg (trend + coherent) |
| Signal | Live sheep imports from France doubled in 2023 vs 2022 (HS 010410) |
| 2025 structural shift | Sheep/goat pox emergency measures tightened movement rules and prohibited restocking of affected establishments- meaning “imports/restocking as a quick fix” is no longer a safe assumption |
Methodology (minimal math, maximum governance)
I ran a rolling origin backtest (repeatedly re-estimating and forecasting) because the only accuracy that matters is out-of-sample accuracy.
1) Seasonal Naive (SNAIVE) as the strict baseline
For monthly data with seasonality m = 12:
ŷ(t+h | t) = y(t+h−12)If a “smart” model cannot beat SNAIVE, it is not forecasting—it is storytelling.
2) WAPE for operational accuracy
WAPE = Σ|y_t − ŷ_t| / Σ y_tWAPE speaks directly to volume planning: “How big is my error relative to the scale of the system?”
3) Coherence constraint (National must equal the sum of prefectures)
For each month t:
Ŷ_GR(t) = Σ_i ŷ_i(t)4) Reconciliation (one practical implementation)
A simple and effective reconciliation approach is scaling prefecture forecasts so that their sum matches the national forecast:
k_t = Ŷ_GR,direct(t) / Σ_i ŷ_i(t)
ỹ_i(t) = ŷ_i(t) · k_tThis is not “maths for the sake of maths.” This is the difference between a nice chart and a tool that can drive milk-zone decisions without internal contradictions.
Backtest outcome: ETS as “champion”, SNAIVE as “floor”
In the rolling backtest, ETS emerged as the best performer, while SNAIVE remained a strong baseline. The point is not the decimal places; the point is that accuracy is measurable, repeatable, and operationally defensible.
| Model | Role | WAPE (approx.) | Interpretation |
|---|---|---|---|
| ETS (Exponential Smoothing) | Champion | ~2.30% | Best “workhorse” performance on seasonal noisy series |
| SNAIVE (Seasonal Naive) | Floor / baseline | ~2.59% | “Last year’s same month”—if you can’t beat this, stop |
2026 is not a single number. It is a corridor.
For 2026, I keep a corridor rather than a point estimate:
- Conservative / floor: ~734m kg (SNAIVE-driven)
- Trend + coherent: ~768m kg (ETS-driven with reconciliation)
Operationally, that corridor is the difference between “business as usual” and a material shift in milk collection pressure, plant scheduling, standardisation constraints, feta cheese production.. and working capital.
The “France signal”: live sheep imports as a proxy for capital entering the system
When I want to understand whether the sector injected livestock capital into the system, I do not look only at milk volume. I look at a leading indicator: imports of live sheep (HS 010410). France provides a clean signal.
According to WITS/UN Comtrade-derived trade statistics, Greece imported the following from France:
| Year | Product | Partner | Quantity (Items) | Trade value (USD ‘000) |
|---|---|---|---|---|
| 2022 | HS 010410 (Live sheep) | France | 7,038 | 1,696.21 |
| 2023 | HS 010410 (Live sheep) | France | 14,224 | 2,001.96 |
This is a doubling in animals (+102% in items). I read it as “capital entering the system” with a 12–36 month lag. In other words: part of what feels like a “good period” today is not only market conditions- it is the delayed effect of capital replenishment.
Sources: WITS country import tables for Greece, HS 010410 (2022 and 2023).
2025 changed the restocking assumption: sheep/goat pox emergency measures and movement constraints
Here is the structural shift that matters more than any decimal in a forecast: in 2025, with sheep/goat pox pressure, the system learned that “if we have a livestock gap, we will import/restock quickly” is not a safe operational assumption.
EU emergency measures for sheep and goat pox in Greece (amending Implementing Decision (EU) 2024/2207) explicitly tightened restrictions related to movements of sheep and goats within restricted zones, imposed conditions for movements (authorisations, testing requirements, direct-to-slaughter logic in some cases), and critically introduced a prohibition on restocking affected establishments. This is not a theoretical policy note: it directly impacts how quickly (or whether) the system can “refill” production capacity after disruption.
In Commission Implementing Decision (EU) 2025/1287, the recital and amendments include “additional restrictions in relation to movements of sheep and goats… and suspend the restocking of previously affected establishments,” and the decision adds that “restocking of affected establishments with sheep and goats shall be prohibited.”
Operationally, that means: even if supply-side economics look supportive, the risk envelope changes because the usual shock absorber, rapid restocking/import-driven recovery, can be partially or fully constrained during outbreaks.
Why I frame 2027 as a cliff risk
If 2023 was a “head injection” year (at least partially), and 2025 reminded us that restocking and movements can be constrained under animal-health pressure, then 2027 becomes a natural horizon for risk pricing. Not because we know the exact policy shape, but because the sector enters the transition logic of a new CAP cycle after 2027, with active proposals and institutional debate already on the table.
In that context, I treat 2026 as a volume corridor to plan for, and I treat 2027+ as the point at which resilience (biosecurity discipline, contracting depth, geographic risk management, and operational readiness) will be tested harder than the sector expects.
What I take away
- Forecasting is not “a number.” It is a governance system: accuracy + coherence + decision usability.
- Geography matters. “National stability” can hide violent internal redistribution.
- Livestock imports signal matters. Live sheep imports from France doubled in 2023 vs 2022, a transparent proxy that animals entered the system and are in high production in 2026.
- 2025 changed the rules of recovery. Under sheep/goat pox emergency measures, movements are constrained, and restocking of affected establishments may be prohibited, so “imports/restocking as a quick fix” is not guaranteed.