Methodology

What this site is

This site rates Canadian dragon boat crews from the race results that regatta organizers, timing companies and governing bodies publish: 337 regattas from 2000 to 2026, 100,595 crew results in all. A rating is an estimate of how fast a crew is compared with the other crews of the same boat size, built from the finish order and the time gaps of every race the crew has paddled in Canada. It is not a measure of a crew’s worth or effort, it never compares a standard boat (20 paddlers) with a small boat (10 paddlers), and it says less about a crew the fewer races it has.

It is written for paddlers, coaches and organizers. The rankings page lists ranked crews by rating; a crew’s page shows its rating over time and every race behind it. The rest of this page explains how the numbers are made and how far to trust them.

Which races count

Only races at regattas hosted in Canada are rated, and only over 200, 500 and 2000 m. A standard boat (20 paddlers) and a small boat (10 paddlers) are rated separately. Fun and novelty races, relays and pursuits, standings tables that are not a race, and races that failed a consistency check (an empty result, places out of time order, places that restart halfway down the list) are left out. When a results sheet prints no distance, the winning time decides it: 8,133 rated races take their distance that way and 6,971 from the sheet.

Of the 18,625 races in the archive, 15,104 are rated, holding 81,429 crew results. The rest are left out by these rules:

The four rating views

Every rated race counts in two views: the view for its distance and the all-distances view of its boat size. So each boat size has 4 views, 200, 500 and 2000 m and all distances, 8 in total. A crew has a separate rating in every view it has raced in, and ratings are never compared across views: a 200 m dash and a 2000 m race reward different crews.

The headline rankings use the Standard boat (20 paddlers), all distances view. Most crews race only a few times a season, so pooling the distances gives each rating more races to stand on, and it costs nothing in prediction: on the 2022 and 2023 seasons its predictions were 12% better for 200 m races than the 200 m view’s, 4% better for 500 m races than the 500 m view’s, 23% better for 2000 m races than the 2000 m view’s (by the log-loss measure explained below).

How a rating is calculated

The method is a version of Glicko-2, a rating system from chess that tracks not only a rating but how sure it is of it, adapted here to races with many boats and to time gaps. In plain words:

• Everyone starts equal. A crew’s first rating is the average of the crews already rated, with the widest possible uncertainty, so its first few races set its rating almost on their own.
• Each race moves you toward or away from the crews around you. A race is treated as a set of head-to-head matches, one against every other crew in it, and a race counts about as much as one head-to-head however many boats are in it. Beating a crew moves you up and it down; losing does the reverse.
• Surprises move more. Beating a crew rated far above you moves both ratings a lot; beating a crew rated far below you moves them very little.
• The margin inside a race counts. Finishing 1.1% of the winning time ahead of a crew counts as about three-quarters of a win, 2.9% or more as a full win, and a photo finish almost as a tie. Only gaps inside the same race are used: a time is never compared with a time from another race, venue or distance.
• Uncertainty shrinks with races and grows between seasons. Each race narrows a crew’s ±. Over the winter every crew’s ± widens again (by 600 points for a crew whose ± had shrunk to nothing, less for the rest), never beyond a newcomer’s ± 2,500, so a crew that comes back after years away is treated almost like a new one.
• Races in the same slot are one step. Heats that run at the same time are all scored from the ratings before them, so the order inside a slot doesn’t matter.

A worked example

Why did my rating go down after a win?

Because a rating is not moved by who won. It is moved by how each result compares with what the ratings expected before the start. Two things follow.

• A narrow win over lower-rated crews can cost points. A crew rated well above the others in its race is expected to beat them clearly. The margin counts, as described above, so a win by a small gap counts as less than a full win. When that is less than the ratings expected, the winner’s rating goes down.
• A win in one race can sit next to losses at the same regatta. Every rated race of a regatta moves the rating, heats included, and the chart on a crew’s page draws one point per regatta: the rating after the crew’s last race there. A crew that wins its final after finishing behind lower-rated crews earlier in the day can leave the regatta with a lower rating than it arrived with.

A fall after a win does not mean the crew raced badly. It means the ratings had expected a little more than the result showed.

What the number and the bar mean

Ratings are calculated on a very wide scale. What the site shows is that rating moved onto a friendlier scale: the shown rating is 1,500 plus 0.5 times the calculated rating’s distance from 1500 (in the model’s own notation, display = 1500 + 0.5 * (rating - 1500)). The change is the same for every view, so order and the meaning of a gap are unchanged. The lowest rating among ranked crews today is 311; there is no fixed top.

The gap between two crews’ ratings sets the chance that one finishes ahead of the other. For two crews with a typical uncertainty, shown as ± 304:

The ± after a rating, and the bar beside it on the rankings page, is the range the model considers plausible for the crew’s true strength: the rating plus or minus twice the model’s uncertainty about it. A wide bar means few recent races; a narrow bar means many. The wider the two crews’ bars, the closer to even the chance between them.

Who gets ranked

A rank is given only to crews whose rating rests on enough recent racing (ranked crews are ordered by rating alone).

Ranks are within a gender class only: Mixed, Women or Open. The end of this section says why there is no overall ranking.

A crew is ranked when all of these hold:

• Canadian crews only. Crews whose home is in the United States are rated, because their races carry information about Canadian crews, but they are not listed in the rankings.
• Active. The crew raced in 2026. A crew whose last race was before 2024 is inactive and keeps its rating without a rank.
• The Nationals floor. A crew is ranked when it raced at least as much as the least-active crews at that season’s Canadian National Championships. The floor sits at the point below which only 10% of the Nationals crews raced, rounded down, and never below 2 regattas and 6 races, counting all a crew’s rated races at any distance that season. A crew that raced this season qualifies by meeting this season’s floor, or by having met the floor of one of the 2 seasons before it. Seasons without a Nationals edition in the archive inherit the last earlier floor.
• Enough races in the view. At least 3 rated races in the view itself.
• Provisional. Until a crew has 3 rated races in a view or its ± has narrowed below 790 (790 on the calculated scale), it is provisional: listed, not ranked.
• Too uncertain to rate. A crew whose ± is 1,500 or more (1,500 on the calculated scale) has no shown rating at all; at that width the number says little about its next results. Most such crews have not raced in years.

The floors for the seasons that count today:

• In 2024, ranked crews raced at least 2 regattas and 6 races, inherited from 2023, the last earlier season with Nationals results.
• In 2025, ranked crews raced at least 2 regattas and 6 races, the same as the least-active crews at that season’s Nationals.
• In 2026, ranked crews raced at least 2 regattas and 6 races, inherited from 2025, the last earlier season with Nationals results.

Crews that are active but miss the floor are listed in the Emerging section under the rankings, with their rating and race count but no rank. In the headline view today, 335 crews are ranked and 1,457 are emerging.

Why isn’t there one overall ranking?

Because the results do not support one. Crews of different classes rarely race each other: open, women’s and mixed crews are tied to one another only through the few crew names that enter races of two classes, which is too thin a link to say that an open crew ranks above or below a mixed one. Ratings are comparable within a class, so each ranked crew has a rank among the Mixed, the Women’s or the Open crews of a view, and the rankings page has a tab per class. The All tab lists every ranked crew by rating with its class beside its name, for browsing, without a rank.

How crews are identified

Results are kept at crew level only: the crew name and club as printed, never paddlers’ names, even where a results sheet prints a roster.

A crew is the same crew across regattas when its printed names match after cleaning: letter case, accents and punctuation are ignored, club acronyms are expanded from the club column where a sheet prints one, and division labels at the end of a name are set aside. Two names that appear in the same race are never merged. Names that match only approximately are merged only when club, province and seasons agree, and never when a crew colour, a number or an age or gender class tells them apart. 11,987 distinct printed names resolve to 11,697 crews: 11,287 by an exact match after cleaning, 148 by an approximate match, 405 by a recorded decision, and 171 kept apart on purpose (entry codes and placeholder names that only make sense inside one regatta).

The province shown for a crew is its club’s when the club’s home is known (from the club’s own name, or because all of its crews race mostly in one province); otherwise it is the province where the crew has most results, and the crew’s page says “races mostly in”. The province filter and the Canadian-crews rule use this same value.

A wrong merge or split moves two crews’ ratings. To report one, send the regatta, the race number and both crew names as printed (see “Reporting an error” below). Every merge decision is recorded, so a correction takes effect at the next build.

How good is it?

The settings were chosen and the method checked on seasons the model had not seen. Ratings were built from results up to 2021; the settings were chosen on the 2022 and 2023 seasons; then, once, the frozen method was scored on the 2024 to 2026 seasons.

Two measures, each taken over every pair of crews in every race of the held-out seasons: log-loss rewards a method for giving the crew that finished ahead a high chance and punishes confident mistakes (lower is better; giving every pair even odds scores 0.693); accuracy is how often the crew the method favoured did finish ahead. On the 2022 and 2023 seasons, counting only races where every crew already had 3 rated races and belonged to the main connected group of crews:

Elo is the rating system used in chess, which counts only wins and losses; “average finish” ranks crews by their average place so far; “coin flip” gives every pair even odds.

On the 2024 to 2026 seasons, scored once in the headline view: log-loss 0.479 and accuracy 77% over 2,839 races (74,462 pairs of crews); on the 1,149 races between established crews in the main group, log-loss 0.456. These figures were measured before the latest corrections to which printed names belong to one crew (see Changes below). The settings were not chosen again and the held-out seasons were not scored again, so the ratings shown today rest on slightly different results than the ones these figures describe.

Calibration asks whether the chances the model gives come true as often as it says. On the 2024 to 2026 seasons, when the model gave a crew about a 75% chance of finishing ahead, that crew finished ahead 80% of the time; when it gave about 25%, the crew finished ahead 20% of the time. Its chances are a little cautious: real outcomes are more one-sided than it says, so a shown chance well above even is, if anything, an understatement.

Known weaknesses

• Sparse provinces. Saskatchewan has 139 rated races in the archive (none in the checked seasons); Nova Scotia has 374 rated races in the archive (none in the checked seasons); Alberta has 739 rated races in the archive (380 in the checked seasons). Crews that race only at home are rated against a small field, and their ratings connect to the rest of the country only through the few crews that travel.
• Crews from different provinces are harder to compare. Log-loss 0.488 on races with crews from more than one province, against 0.477 when every crew is from the same province.
• Cold starts. Races with a crew that had fewer than 3 rated races: log-loss 0.495, against 0.456 between established crews. Most crews race a handful of times a year, so many ratings stay provisional.
• Cautious chances. As the calibration sentence above says, the chances shown are a little too even; a method that sharpens them is on the list for later.
• Thin views. Too few races for a steady rating or a ranking: Small boat (10 paddlers), 200 m: 206 rated races, 4 ranked crews; Small boat (10 paddlers), 2000 m: 38 rated races, 3 ranked crews; Small boat (10 paddlers), 500 m: 217 rated races, 4 ranked crews; Small boat (10 paddlers), all distances: 461 rated races, 4 ranked crews.
• Crew identity. A wrong merge or split of two names moves both ratings (see above).

Where the data comes from

Every result comes from an original source: the regatta organizer, its timing company, a governing body or a club. The raw files are kept as fetched, so a parse can be redone, and nothing is typed in by hand. Pages that only list events by name and date are used to find regattas, never as results.

The archive holds 337 regattas, 18,625 races and 100,595 crew results, of which 81,429 are rated. Since 2015, 99% of results belong to one connected group of crews that have raced each other directly or through others, so their ratings sit on one scale. By province, the share of crews since 2015 that belong to that group:

• Nova Scotia: 100% of its 103 crews
• Ontario: 97% of its 2,264 crews
• British Columbia: 96% of its 1,472 crews
• Alberta: 95% of its 368 crews
• Quebec: 91% of its 1,215 crews
• Saskatchewan: 64% of its 95 crews

The Behind the data page shows the full funnel from fetched files to rated races, coverage by province and year, every consistency check and the assumptions register: the values the data relies on that no source prints, each with its basis.

Reporting an error

There is no form yet; one comes in a later phase. Until then, send reports to results@dragonboatratings.ca. A report says which regatta, which race number and which crew as printed, and what you expected instead; a link to the organizer’s results page settles most cases. Corrections are recorded with their evidence and applied at the next build.

Changes

Methodology changes by date, from the project’s decision log:

• 2026-11-13: Crews of clubs whose home is in the United States (the place is in the club's name, or the results print it) are no longer listed as Canadian crews because they raced mostly at one Canadian regatta; a club page counts Canadian-home crews only and names US crews apart. A club's province is no longer stated on the evidence of two crews that only share the beginning of a name. No rating or rank changed. (D131)
• 2026-11-12: A crew is now shown under its club's province when the club's home is known, and otherwise as "races mostly in" the province where it has most results. Before, every crew was shown where it had raced most, so a travelling crew of an Edmonton club could appear under British Columbia. No rating or rank changed. (D121)
• 2026-11-12: The pages now describe one rule in one way: a crew that raced in a season is ranked on that season's floor or on the floor of one of the two seasons before it. The home page's earlier sentence could be read as three earlier seasons; the rule itself did not change. (D120)
• 2026-11-12: For the crews at the top of each class, results printed under a second name (a class word or a club prefix more or less, a former club) were joined to the crew; the settings were kept and the held-out seasons were not scored again. (D119)
• 2026-11-12: Ranks now count within a gender class only (Mixed, Women, Open); the All tab lists every ranked crew by rating without a rank, because crews of different classes rarely race each other. (D110)
• 2026-11-11: Settings re-tuned after about 10,900 newly read results; in the headline view a narrow win now counts as slightly more of a win; the method was scored once more on the held-out seasons, per view. (D108)
• 2026-11-11: A club that enters one crew name in two or more of the women's, open and mixed classes at one regatta now counts as one crew per class; a lone "F" in a race title is read as a division letter, not as a women's race. (D107)
• 2026-11-10: Small-boat races are now recognised from the word in the race title and rated in the small-boat views; a race in which small and standard boats start together is left out. (D102)
• 2026-11-10: Every race that is left out of the ratings now says which rule left it out, in plain words, and times are printed exactly as the results sheet printed them. (D92)
• 2026-11-09: Settings re-tuned on the current data, view by view; the headline view's season widening and margin scale changed; the method was scored once more on the held-out seasons, per view, and the model card was rewritten for the views. (D86)
• 2026-11-09: The Nationals floor counts a crew's regattas and races at any distance, and a view also needs three rated races in it; where a sheet prints no distance, the winning time decides it, with checked time bands. (D79)
• 2026-11-09: Only regattas hosted in Canada and only the three standard distances are rated; four views per boat size (one per distance and all distances) replace the sprint and long pools; only Canadian crews are ranked; the Nationals-attendance floor decides who gets a rank; ratings are shown on a friendlier scale. (D78)
• 2026-11-08: Crews stay ordered by rating alone, not rating minus uncertainty, because that order predicted the next season best; a crew whose uncertainty is very wide has no shown rating and no rank. (D75)
• 2026-11-07: First published ratings: one rating per crew and view, crews ordered by rating; the chance of one crew finishing ahead of another is computed from the two ratings and their uncertainties. (D74)
• 2026-11-07: Time gaps inside a race now count: a clear win moves ratings more than a photo finish, and a time is never compared with one from another race. This margin-aware method became the production model after scoring best on the held-out seasons. (D73)
• 2026-11-07: Ratings now carry an uncertainty that shrinks with every race and widens between seasons; a crew is provisional until it has three rated races or its uncertainty is narrow enough; settings are chosen on past seasons only. (D70)
• 2026-11-07: Which races count: standings tables, novelty, relay and pursuit races, and races that fail a consistency check are left out; the finish order comes from the printed places, or from the times when places are missing; races that run at the same time are scored together. (D67)