About Rankquant

The rating-inflation problem

Open Amazon. Every product is 4.4. Open Yelp. Every restaurant is 4.2. Open Wine Spectator. Every "classic" wine is 92–97. Open Booking.com. Every hotel is 8.4. Signal is dead across every review surface on the internet.

If "excellent" applies to everything, it applies to nothing.

We think reviews should help you make real buying decisions. Averages don't.

Our fix: a normalized percentile, not another average

Rankquant takes raw ratings from the usual sources — professional critics, verified-purchase aggregators, enthusiast platforms — and runs them through a four-step statistical pipeline that's standard in professional practice but has never been published as the basis of a consumer review site:

1. Per-reviewer z-score normalization. For every reviewer we compute their personal mean μ_u and standard deviation σ_u, then convert each of their ratings to z_u,i = (r_u,i − μ_u) / σ_u. Personal scale washes out; what's left is each reviewer's opinion of relative quality.
2. One unweighted aggregate per product. We average those z-scores across every qualifying reviewer. Every reviewer counts equally — no reviewer and no source is weighted more heavily than another, there is no credibility multiplier and no per-source weight table anywhere in the pipeline.
3. Empirical-CDF percentile.Every product's mean z-score is ranked against every other product's mean z-score in the database and expressed as a 0–100 global percentile. A 90 means top 10%. The in-cohort percentile is that same z-score re-ranked within same-category, ±20%-price peers — a smaller reference set, not a different calculation.
4. A confidence correction alongside it.The AI-adjusted percentile takes statistical uncertainty into account: a product keeps the share n / (n + 53) of its measured distance from the corpus average, so a 4-reviewer +2.1 mean is pulled almost all the way back to the average while an 80-reviewer +1.6 mean keeps most of its distance. Small samples can't fake-rank. This adjustment lives in the AI-adjusted figure alone: the global and in-cohort percentiles report the measurement as made, on every vertical.

Why we start with wines

Wine criticism has the longest-running quantitative scoring culture in any consumer category — Robert Parker's 100-point scale has been around since 1978. It's also the most inflated: the effective range is 85–100, compressing what should be 50 points into 15. Forty years of grade inflation have destroyed the signal.

If our methodology can cut through 40 years of wine-score inflation, it can cut through anything. Wines are the proof-of-concept. Movies, books, Amazon products, hotels, and restaurants are next. Eventually every major consumer category.

Why this is better

Consider buying noise-cancelling headphones. Sony WH-1000XM6 averages 4.5 on Amazon. Bose QuietComfort Ultra averages 4.5 on Amazon. Tie. Useless.

On Rankquant:

• Sony WH-1000XM6 → global percentile 97 (Ẑ = +1.75, 142 qualifying reviewers)
• Bose QuietComfort Ultra → global percentile 86 (Ẑ = +1.05, 128 qualifying reviewers)

Now you have a decision. Both are great products, but one is measurably better relative to the category. That's the information buried in the raw data that averaging throws away.

Now consider wines. 2019 Louis Jadot Bourgogne Chardonnay scores 89 on Wine Spectator. So do 400 other wines. Useless. On Rankquant it sits at global 62 and in-cohort 91within its same-category $15–$30 cohort — not a standout globally, but a best-in-price-class pick. That's a real buying signal.

Why you should trust our logic

Rankquant is led by Ryan Siegal, whose fifteen-year quantitative-finance career is the source of the methodology. Ryan founded a multi-strategy hedge fund (14% average annual return, $12M AUM), traded event-driven and merger-arbitrage strategies at Alpine Global with 25–35% annual returns, and now applies Python and AI/ML to research and systematic process improvement at Relentless Upside Consulting. Washington University in St. Louis alum. Consulting statisticians Ben, Josh, and Yang contribute STEM-trained input on distribution analysis, normalization design, and outlier detection.

The methodology behind Rankquant — z-scoring, confidence intervals, shrinkage estimation — is the same statistical toolkit that runs trading systems. None of the math is novel. What's different is publishing it openly, applying it rigorously, and committing to constants up front so the output is reproducible. We don't claim novel math. We claim applied math — taking methods that work in finance and psychometrics every day, and using them on review data where the field has been making do with raw averages for thirty years.

Everything we do is documented at /methodology. Every review page shows the math step-by-step: raw scores per source, the formula, the intermediate values, the final score. If you disagree with the output, you can check our work. The normalization is specified in full — admission rules, the shrinkage constant, the percentile mapping — at /methodology, so any percentile we publish can be recomputed and contested.

Independence and disclosure

Product pages, book pages, the film/TV catalog panel and the wine pages whose exact bottle Amazon sells carry Amazon Associates tracked links, which may earn us a commission at no additional cost to you. Every other wine link, and the hotel and Cruise Critic links, carry no tracking parameter and earn us nothing. Each category has one fixed outbound destination, built from the record's own identifiers after the score is computed — there is no price-and-commission routing. The full description is at /methodology#affiliate. No outbound link affects the normalized score. Brands cannot buy a higher ranking.

How to use this site

If you're buying something right now:

1. Go to the category (/wines; later /movies, /books, /headphones, etc.).
2. Look at the top 5 by normalized score.
3. Read the methodology-transparent breakdown on the product page.
4. Click through to the retailer we've selected as best-for-you.

If you're a brand or researcher: per-category distributions will be published monthly at /data. The normalization procedure is specified in full at /methodology.

If you have feedback:every review page has a "report a correction" link. We take methodology criticism seriously. If you spot a statistical error we'll fix it, publicly, and attribute you.

Our editorial policy

• No product is listed without at least 50 real reviews across at least 2 sources.
• Sponsored content is never re-ranked. Brand partnerships are labeled clearly.
• Categories are defined by our editorial team; we do not split or merge categories to benefit any specific product.
• Our top-pick logic is reproducible: publish the same raw inputs on the same date, get the same output.
• Historical scores stay visible as data evolves.

Based in New York. Content is verified and re-checked monthly.