The Question

European grocery retailers earn 25-30% margins on private label products, compared to roughly 1% on national brands. So if you're a supermarket deciding where to invest in your own-brand range, the question isn't really whether to do it, but where. Which product categories have the biggest gap between what consumers are looking for (healthier options) and what's actually on the shelves?

I built a pipeline to answer this using open data. The core dataset is Open Food Facts, a crowd-sourced database of 2.57 million food products across 27 EU countries, enriched with pricing scraped from Mercadona (Spain) and Albert Heijn (Netherlands). The idea is to quantify, per product category, how unhealthy the current offerings are and how little private label has done to fill that gap.

The headline finding: 73% of EU food products score Nutri-Score C or worse, while private label penetration among healthy (A/B) alternatives sits at about 10%.

Data Collection

Open Food Facts is large but messy. Product names are nested arrays of structs, nutriments are stored the same way, and country tags come back as numpy arrays rather than plain lists. About 68% of products are missing their official Nutri-Score grade, so I computed it for 706,000 of those using the published 2023 algorithm. That brought coverage from 32% up to 59%.

For pricing data I scraped two retailers. Mercadona's API is straightforward REST with a three-level category hierarchy, but it doesn't include barcodes, so matching to Open Food Facts required fuzzy name matching (51% match rate via rapidfuzz). Albert Heijn's mobile API needs anonymous auth with a clientId: "appie" header, but it does return brand names and Nutri-Score grades directly. Match rate there was 41%.

Private label detection is its own problem. I compiled a list of 80+ known PL brand names across 15 European retailers (Hacendado, AH Biologisch, etc.), but that only catches products where the brand field is actually populated. To fill the gaps I trained a TF-IDF character n-gram classifier on the labelled supermarket data. It achieves a cross-validated F1 of 0.996, which makes sense given how distinctive PL product naming patterns are.

The Nutritional Landscape

Sorting 45 food categories by the proportion of products scoring C, D, or E on Nutri-Score gives you a pretty stark picture. Categories like Taralli, Sweet Pies, and Bread Coverings are almost entirely unhealthy. Even broad categories like Snacks (96% CDE) and Breakfast (98% CDE) have almost no healthy products on offer.

But a category being unhealthy isn't by itself an opportunity. What matters is whether private label has already moved into the healthy end of that category. The quadrant chart below puts these two dimensions together. The x-axis is the percentage of products scoring C/D/E (unhealthiness), and the y-axis is private label penetration among A/B products. Bottom-right is where you want to look: lots of unhealthy products, very little healthy private label.

I combine these into a single "nutritional gap" metric: the fraction of products scoring CDE, multiplied by one minus the PL penetration at healthy grades. Breakfast leads at 0.935, followed by Snacks (0.879) and Desserts (0.839).

What's Making Them Unhealthy?

Looking at the median nutrient profile of CDE products in each high-opportunity category tells you what a reformulation effort would actually need to target. The heatmap below shows the six nutrients that feed into the Nutri-Score algorithm, with values normalised so the colour scale is comparable across nutrients.

Breakfast products are overwhelmingly a sugar problem (55g/100g median). Fats & Oils is, unsurprisingly, saturated fat (28g/100g). Snacks have high sugar and high fat, which makes reformulation harder because you're fighting on two fronts.

To quantify how feasible reformulation actually is, I ran a binary search on each category's median nutrient profile: what's the minimum reduction in sugar, salt, or saturated fat needed to push the Nutri-Score from its current grade to a B?

Crepes & Galettes needs a 37% sugar reduction and 56% sat fat reduction, which is aggressive but within range for food science. Taralli needs about 40% salt reduction. But Breakfast needs an 84% sugar reduction, which essentially means making a different product. And Fats & Oils needs 96% sat fat reduction, which is physically impossible while remaining oil.

Composite Scoring

The nutritional gap alone doesn't tell you where to invest. You also want to know: is the category big enough to matter? Is it dominated by a few brands (hard to enter) or fragmented (easy)? Can the products actually be reformulated? What kind of price premium can PL capture?

I built a six-component composite score, with each component normalised to [0, 1]:

opportunity = 0.25 × nutritional_gap       # how big is the health gap?
            + 0.15 × (1 - HHI)              # brand fragmentation (lower = easier entry)
            + 0.15 × log(category_size)      # market size
            + 0.15 × reformulation_ease      # can it actually be made healthier?
            + 0.15 × (1 - PL_saturation)     # room for new PL entrants
            + 0.15 × price_gap_margin        # brand-to-PL price premium

Nearly all categories have low brand concentration (HHI below 0.05), meaning the markets are highly fragmented. That's good news for a new entrant. The median PL discount across all categories is 22.4% below branded equivalents, which gives decent margin headroom.

Crepes & Galettes comes out on top (0.689), mainly because its reformulation path is the most realistic. Condiments & Sauces (0.672) benefits from being a large market with a clear salt-reduction path. Breakfast (0.663) has the highest raw nutritional gap but gets penalised on reformulation feasibility. Snacks (0.657) is the biggest market by far at 119,000 products.

Sensitivity Analysis

Any composite index raises the question of how sensitive the rankings are to the choice of weights. I ran 1,000 Monte Carlo simulations drawing weights from a Dirichlet distribution, then checked how much each category's rank shifts under different weighting schemes.

The top four (Crepes & Galettes, Condiments & Sauces, Breakfast, Snacks) stay in the top five across almost all scenarios. Snacks jumps to 1st under a size-heavy weighting (it's the largest market). Taralli is the most volatile, jumping between 1st and 37th depending on whether you prioritise brand fragmentation or reformulation ease.

Predictive Modelling

As a supplementary analysis, I trained a gradient boosted classifier to predict which products become PL "leaders" (top 3 by scan count in each category). The model uses 17 features (9 nutrients, Nutri-Score grade, NOVA classification, and 5 label flags) and achieves a cross-validated AUC of 0.653. That's modest, which isn't surprising given that scan counts are a noisy proxy for actual sales.

The useful output here isn't prediction accuracy, it's the feature importances. Energy, sugars, salt, and protein are the top predictors, and PL leaders tend to have slightly lower sugar (2.8 vs 3.6g/100g) than the category average. The direction is consistent with the gap analysis: there's a demonstrated consumer preference for marginally healthier products within PL ranges.

Dashboard and Pipeline

The full pipeline runs as a sequence of CLI scripts: download the 4.4 GB OFF Parquet from Hugging Face, run the supermarket scrapers, then build the cleaned and joined dataset. The analysis notebooks go through EDA, category landscape, nutritional gaps, opportunity scoring, and predictive modelling in order.

Everything feeds into a Streamlit dashboard with four tabs: category landscape (treemap and HHI bar chart), nutritional gaps (quadrant scatter with interactive filtering), opportunity ranking (sortable composite scores with component breakdown), and data quality (coverage stats and limitations). It loads from pre-computed sample Parquets, so there's no heavy processing at runtime.

There are 51 tests covering Nutri-Score computation, EAN and fuzzy joining, gap analysis, opportunity scoring, and the brand classifier. All run on synthetic data, no downloads needed.