I wanted a project that would force me to do real MLOps, not just train a notebook model and call it done: a shared feature pipeline between training and serving, a model registry, scheduled retraining, and honest evaluation against a holdout instead of a cherry-picked demo. I also wanted an excuse to use AWS's own ML services hands-on rather than read about them. Wine prices turned out to be a good problem for that — there's a real public dataset behind them, and the question "is this about to get more expensive" has an obvious, checkable answer a quarter later.
The result is Cellar Index, live at thepriceofwine.com. It tracks about 31,000 wines and forecasts which ones are likely to move in price, whether today's price is fair for what the wine is, and how the next vintage is likely to be priced against the current one.
The data: 508,000 rows from 40 PDFs
Pennsylvania is a control state — the Pennsylvania Liquor Control Board is the retailer, and Act 39 of 2016 requires it to publish the full retail price list every quarter. That gives ten years of real, official shelf prices: 40 quarterly reports, each with 12,000–15,000 lines of code, description, size, regular price, and any sale or clearance price. No scraping fragile retailer pages, no synthetic placeholder data — just a government PDF that changes format on you once in a decade.
Parsing it was most of the early work. The report layout changed in October 2023, names wrap onto extra lines, and there's no category column, so keyword rules split wine from spirits and accessories and assign a style. The harder decision was what counts as "one wine." I settled on: one vintage of one product at one bottle size. Caymus 2019 and Caymus 2020 are different wines with separate price histories. A union-find join ties rows together within a vintage, using the PLCB code where it survives the 2023 rewrite and a normalized name otherwise, and vintages of the same product are grouped into a family so a wine's page can show its other vintages. About 30% of rows print no vintage at all — some are genuinely non-vintage (Champagne NV), others are everyday wines the PLCB just doesn't record a vintage for — and those are tracked as their own "NV" series rather than forced into a vintage they don't have.
Three models, one real question
List prices are sticky. About 3% of wines change price in a given quarter, so "predict no change" is right almost all the time and useless as a product. The actual question is which wines are about to move. Cellar Index answers that with a price-change classifier — P(higher), P(lower), and the typical size of the move, at one to four quarters out — calibrated with Platt scaling and conformalized quantile regression for the price ranges. On the last four holdout quarters it never saw during training, it hits 0.89 ROC AUC for next-quarter rises and 0.84 at a year out, and flagged wines are 3–6× more likely to actually move than an average wine.
Two more models sit alongside it. A next-vintage model predicts how a wine's next vintage will be priced against the current one, trained on about 4,900 past vintage transitions; it beats a flat "same price" guess in every one of four rolling test years, by 4–21% in mean absolute error. And a fair-price model — a hedonic pricing model, the standard approach in wine economics — predicts what a wine should cost given what it is: region, grape, classification, style, bottle size, vintage age, the words in its name, and the producer's price level from its other wines, but never the wine's own price. Cross-validation is grouped by wine family, so a wine is always priced by a model that saw none of its vintages. On held-out wines it lands a 16.3% median error and explains 80% of the variation in log price, against a 36.7% baseline from just region-plus-grape medians.
The part I'm most glad I did: an ablation discipline
It would have been easy to throw every feature I could source at these models and ship whatever scored best on one test split. Instead every feature group had to clear the same bar: does it help across multiple rolling holdout years, not just one. Most didn't.
I built a weather feature from NASA POWER reanalysis data — growing degree days, frost days, heatwave days, harvest and season rain, computed as anomalies against each region's 1991–2020 normal — and it correctly captures real events like the April 2021 Champagne and Burgundy frost and Napa's hot 2020. It helps predict price cuts, but it hurts rise ranking, and its effect on next-vintage pricing flips sign between test years. It's kept as data, not used. A USDA California grape-crush supply feature fared better: it lowered next-vintage error in all four rolling test years and is adopted there. A rule-of-thumb drinking-window feature helped the price-change model meaningfully for vintage-dated wines (rise AUC 0.771 → 0.808) and is adopted.
Brand momentum — the share of a producer's other wines raised or cut recently, computed point-in-time so a wine never sees its own future — looked like a clear win on the most recent holdout year (rise ranking 0.838 → 0.851). On an earlier holdout year it made every single comparison worse. The likely cause is a statewide price increase in April 2023 that touched 31% of all wines at once: the model was probably just learning "prices moved together that quarter," not anything about the brand. Grape variety and classification (Grand Cru, Reserva, DOCG) hurt in both models where I tried them, and in the next-vintage model specifically they overfit a training set of only about 4,900 transitions once the high-cardinality grape category was added. None of these four are adopted. All four are still computed and stored, because rejecting a feature isn't the same as deciding it's wrong forever — it means it didn't earn its place yet.
Making training and serving share one brain
The thing I was most deliberate about: there's exactly one feature pipeline, in a wineprice package, and both the training job and the FastAPI service call the same WinePriceModel.forecast() method from it. That rules out the most common way an ML system silently breaks — training and serving drifting apart because someone edited the feature logic in one place and not the other. An automated parity test checks that online predictions match the stored batch forecasts for a sample of wines on every run, specifically to catch that regression.
Each training run writes to an immutable, versioned model registry — local folders in development, S3 in production — with its metrics and artifact path recorded and exactly one version marked active. The API polls for the active version and hot-swaps a newly promoted model within 60 seconds, no restart. Online inference for a single wine, including the what-if scenario endpoint ("what would this forecast look like at a 40% clearance discount"), runs in about 110ms.
It's deployed on AWS with CDK: one EC2 instance behind CloudFront, Postgres on the same box, no SSH — admin access is through SSM Session Manager only. A quarterly EventBridge schedule, timed a week after each new PLCB price list, triggers an SSM command that ingests the new data and kicks off retraining as a SageMaker Processing job (the account's training-job quota is 0, so Processing jobs stand in). Trained artifacts land in the S3 registry and the API picks them up automatically. A Bedrock call to Amazon Nova Lite turns each wine's forecast, fair price, and drinking window into a plain-English explanation on click, cached and rate-limited so it stays cheap. The whole thing — compute, storage, and quarterly retraining — runs for about $17 a month.
Model tuning still to come
A few things are on the list for the next round, all of them problems the evaluation already surfaced rather than guesses about what might help:
- Fixing the calibration drift. At three to four quarters out, the model predicts a 4.7% rise rate against an actual rate of 2.6%, because price rises slowed down after 2024 and the recent calibration window still reflects the faster pace before that. A flat recency window isn't enough here — the fix is probably a calibration that's explicitly aware of the trend, not just the last few quarters, or a decay term that down-weights older, faster-moving periods more aggressively.
- Making brand momentum survive a statewide shock. The feature's instability traces to one event — the April 2023 increase — swamping the brand-level signal in whichever holdout year contains it. Before trying brand momentum again, I'd detect statewide-shock quarters explicitly (a quarter where price-change rate is several standard deviations above normal) and either exclude them from the brand-momentum calculation or let the model see the shock as its own feature, so brand-specific moves and market-wide moves stop getting confused for each other.
- Giving grape and classification a fairer shot. These overfit hardest in the next-vintage model, which only has about 4,900 training transitions to work with — not because the signal is necessarily fake, but because raw high-cardinality categories are too expensive to fit on that little data. The fair-price model has 31,000 wines to learn from instead, so a hierarchical or shrinkage encoding (rather than one-hot or raw categorical) might let these attributes earn their place there even though they don't in the next-vintage model.
- A reputation signal for the fair-price model. Its clearest blind spot is prestige the features can't see: a famous producer's entry-level wine (Mouton's Aile d'Argent) can look like a bargain, and icons like Grange or Opus One show large premiums the model reads as overpricing. The features it has — region, grape, the producer's own price level — can't distinguish a producer's reputation from its price-setting pattern. A critic-score or secondary-market reference feature would probably close most of that gap, but it's a harder data-sourcing problem than anything else on this list, since that data isn't sitting in a government PDF.
None of these are things I'd ship speculatively. The same rule applies as before: they go in only if they clear the rolling-holdout bar the adopted features already cleared, and they come back out if they don't.
See it for yourself
Cellar Index is live at thepriceofwine.com, including a Model page that shows every ablation result above with the real numbers, not just the summary. The API docs are at /api/v1/docs.
Rejecting a feature isn't the same as deciding it's wrong forever — it means it didn't earn its place yet.