How the Mac is Won

The Chicago Yacht Club's Race to Mackinac runs 289 nautical miles from Navy Pier to Mackinac Island. Bayview starts off Port Huron and sends divisions onto two courses: 204 miles up the Michigan shore, or 259 miles around Cove Island and back through the Straits. All three finish in the same narrow water below Mackinac Island.

Three courses, two starting waters

Chicago crosses Lake Michigan from Navy Pier. Bayview's two courses start within about an hour over the same water, giving us two course geometries in nearly the same weather.

One course can be two different races

Across the thirteen Chicago editions with lake-scale wind, the racing fleet's share of hours unable to lay Mackinac ranged from 97 % in 2018 to 1 % in 2024. In 2024 the fleet ran for 80 % of its hours. Six years earlier it beat for almost the entire race.

That range is why later comparisons are checked by wind regime rather than treating “the Mac” as one weather object.

Retirements measure a different kind of difficulty

Across forty tracked fleets the median retirement rate is 7.0 %. The tail is much larger: 47.2 % on the 2026 Bayview Cove Island course, 42.7 % on its Shore course and 32.4 % in the 2017 Chicago race. High retirement can follow a violent squall, but it can also follow a long beat with no strong convective signature. Section 5 keeps those cases separate.

The archive has two important boundaries. The 2020 Chicago race was not sailed, and Bayview used only the Shore course in 2020 and 2021 because the Canadian border was closed. The 2011 Chicago fleet predates the track archive. Section 5 uses its radar only to describe the weather line; it does not reconstruct or compare the fleet.

A common claim is that the winners win at night. The archive does not show a separate night advantage.

The correlation between a boat's night performance relative to its own daytime performance and corrected place is −0.027 across 414 racing divisions in 37 editions. The edition-clustered 95 % interval is [−0.120, +0.062]. Negative would mean that holding more speed at night went with a better place. The interval includes zero, and the design could reliably detect correlations of about 0.128 or larger. A smaller association remains possible.

Compare each boat with itself

For every track leg, we divide speed over the ground by the speed predicted from a published ORC polar at the local analysed wind angle and strength. This gives an achieved fraction. Night retention is the boat's median achieved fraction at night divided by its own daytime value.

The comparison is within a boat, so a constant error in the reference polar cancels. That cancellation is an assumption: day and night can occupy different wind regimes, where the polar's error may differ. Day and night are assigned at each boat's position using civil twilight rather than a fixed clock.

This is a whole-night comparison, not a test of a particular watch or the hours around 02:00. We detected no additional night-specific association at this resolution. Protecting night performance may still matter; the archive does not locate a distinct winning margin after sunset.

Inside a class, sailing farther usually goes with finishing worse. The pooled Spearman correlation is +0.41 across 434 divisions in 40 tracked fleets, and 39 of 40 fleets are positive. The best third of a class typically sails 4 to 13 fewer miles than the worst third on courses of 204 to 289 miles.

A separate sensitivity reduction over the 37 fleets with wind data and its own eligibility filters gives ρ = +0.434 across 487 divisions, 95 % CI [+0.334, +0.525]. Thinning every track to a common 30-minute cadence gives +0.442 [+0.349, +0.525]. It is not an interval for the 434-division primary estimate, but its sign is stable under the alternative population and cadence. Tracker cadence and missing fixes still make the mileage gaps approximate.

Why the comparison stays inside one class

Across a mixed fleet, faster boats can sail farther because they cover more water per hour. Comparing corrected place inside one division reduces that boat-speed confounding, though rating and configuration differences remain. The result is observational: a boat that is in phase with the shifts may both sail less and move faster. The archive cannot separate the shorter track from the sailing that produced it.

The miles grow when the fleet works harder

The largest gaps occur when fleets sail much more than the published course:

The relationship does not disappear on a beat. The within-division correlations are +0.400 in upwind years (163 divisions), +0.434 in mixed years (167) and +0.392 off the wind (120). Wind regime changes the available mileage more clearly than it changes the correlation.

The exception: 2019 Chicago

The only negative fleet is 2019 Chicago, at ρ = −0.240; the top third sailed 1.2 miles more than the bottom third. We do not know why.

The race began with heavy upwind sailing and waves. The wind later died for many boats, and not every boat finished before the shutdown. Its 5.5-knot fleet-hour median is therefore not a description of the opening conditions or of every boat's passage. I sailed aboard the J/120 Proof; we won our class and sailed a lot of distance.

The edition's distance correlation is not explained by its wind angle, median wind, darkness, squall signature or retirement rate. Those checks rule out simple stories, not every mechanism. A boat-by-boat reconstruction of the shifts would be a separate case study.

What the measurement can establish

Recorded tracks underestimate distance because they draw straight lines between fixes. The archive does not establish the direction of the resulting bias in the best-versus-worst group difference. The sample contains finishers with valid corrected results; including incomplete tracks would create the relationship mechanically. Comparisons are made within editions because tracker cadence differs between years.

The result is an association, not the effect of prescribing a shorter route. It says that fewer miles on the recorded track accompany better corrected places inside a class. Across 72 Transpac and Pacific Cup divisions, the corresponding pooled correlation is +0.221 (+0.239 in the Transpac and +0.194 in the Pacific Cup), about half as strong. A lake course leaves more of its total distance available to the sailor.

A Mackinac is a mesoscale race. Lake breezes, nocturnal jets and convective lines are smaller than a quarter-degree global grid. At this latitude one such cell is roughly 20 by 28 kilometres, while Lake Michigan is about 90 kilometres wide. Four or five cells cannot describe the weather across the lake.

The interactive holds the hour and field fixed and changes only the resolution. The coarse view smears the structures a navigator is trying to use.

The lake-scale archive

The analysis uses hourly 10-metre wind from HRRR at 3 kilometres and RTMA at 2.5 kilometres. HRRR covers races from 2014 onward. RTMA reaches back to 2013 and provides an independent comparison during their overlap. Missing hours remain missing; editions before either archive stay in analyses that do not require wind.

These are analyses of what happened, not forecasts a navigator held during the race. Section 6 separately uses forecast files available by each decision time.

Classify the wind against the mark

A boat beating at 45° to the wind can look like it is reaching if we classify its own course over the ground. Instead, each hour is classified by the angle between the wind and the bearing from the boat to its next mark. Under 60° is upwind, 60° to 120° is reaching, and over 120° is running. Cove Island boats point toward the Cove mark until they round it. Shares are weighted by time so tracker cadence does not give one edition more influence.

The 2018 Chicago fleet beat for nearly the whole race. The 2024 fleet ran for most of it. The classifier is deliberately coarse and does not turn those categories into a description of every boat or every hour. In particular, 2019 began with heavy upwind sailing and waves before a later shutdown pulled its fleet-hour median wind down.

What a squall looks like in an hourly field

A convective line can produce a large direction change and speed jump in the same hour. A front may turn the wind without the same speed spike.

The 2022 Chicago race is the clearest convective signature in the archive. The 2017 Chicago race retired 32.4 % of its fleet but sits near the middle of this ranking, while the 97 %-upwind 2018 race also had heavy retirements without a sharp wind discontinuity. Retirement rate and convective violence measure different kinds of difficult race.

A squall crosses a fleet quickly. A long beat can wear down boats and crews for a day without producing a dramatic radar image. The archive contains both, and retirement rate does not rank them the same way as convective intensity.

What the radar can see

NEXRAD Level II provides raw reflectivity and radial velocity. Its beam rises with distance from the station, while the wind that matters to a boat is near the water.

We use a three-kilometre beam-height limit. Beyond it, reflectivity can still show the middle of a storm but cannot represent what reached the deck. The maps distinguish a valid clear observation from water the radar did not sample.

Five measured nights, and one blank

For each boat, the analysis finds the worst reflectivity over its track and the last clear frame before that peak. This times the arrival of rain, not necessarily the gust front, which can run ahead of the echo.

The 2026 Bayview race had the archive's highest retirement rates, 47.2 % on Cove Island and 42.7 % on Shore, after a severe squall. The wind analysis captures the jump, but the national archive has no Detroit radar files for the relevant days. Gaylord did not cover the fleet's southern position. We therefore leave the radar view blank rather than reconstructing it from a different product.

The 2011 Chicago–Mackinac

On 17–18 July 2011 a severe line crossed the Chicago–Mackinac fleet. WingNuts capsized and two sailors died. This paper does not reconstruct the fleet or turn the event into a performance comparison. The US Sailing independent safety report is the record of the accident and its lessons.

The surviving radar answers one narrower question about the weather. The strongest return intensified from 22.5 to 63.0 dBZ in thirty-six minutes, and the 50 dBZ core advanced at about 25 knots toward 127°. Reflectivity is not wind, so we do not convert it into a gust estimate. A line moving at 25 knots covers fifteen miles in thirty-six minutes; that retrospective travel-time calculation is not a claim about the warning available to a crew.

Holding speed through a line

For each boat, speed retention is its average speed over the ground in the hour after the line arrived divided by its average in the hour before. We use raw speed because the reference polars end below the winds in these squalls.

Negative means that holding more speed went with a better corrected place. Greater retention went with better place in 2022 and 2021. The detected 2016 relationship points in the opposite direction; 2017 and 2024 do not separate. Pooled across 71 racing divisions, the interval includes zero. Five nights cannot establish which conditions produce either pattern.

Later retirees held 0.071 less of their pre-line speed than finishers in this five-night sample. A within-night permutation gives p = 0.006, but resampling the five nights gives an interval of [−0.121, +0.050]. Speed retention is not a reefing decision, and slowing may already reflect damage. The result is an association within these nights, not evidence that pressing on is safer.

Why 2017 looked empty

The 2017 Chicago race retired 32.4 % of its fleet, yet only 32 of 296 boats met 35 dBZ. The radar was operating and covered 88.1 % of the analysis grid; its strongest echo was simply not over the fleet. The archive cannot determine whether the event was a dry downdraft.

Radial velocity does not recover a hidden low-level signal here. The median peak velocity difference was 6.5 m/s in 2017, against 27.5 m/s in 2022, but the usable beam was about two kilometres above the water and Doppler velocity requires scatterers. In 2017, 93 % of fleet samples were below 5 dBZ. The measurement cannot distinguish weak outflow from an absence of particles to measure.

The supportable conclusion is physical and limited: the 2017 event left little reflectivity and only a weak, high-beam velocity signature. A sailor watching the radar colours would have seen an almost empty screen. Radar remains essential, but rain is not the only way a Great Lakes night becomes dangerous.

Start with the newest forecast

The simplest navigator follows the fastest route in the newest forecast. Six hours later, it downloads the next forecast and routes again. We call this the rolling deterministic policy.

The replay starts at 20:00 local from a real boat: the tracked boat nearest the middle of its division's geographic spread. From there, the navigator takes the newest usable HRRR route and sails it for six hours through the hourly HRRR/RTMA analysis. It then routes again from the position it actually reached.

That last detail matters. A new forecast may favor a route that began somewhere else, but the boat cannot move sideways to meet it. The black line in the map is distance already sailed. Move the forecast-update control and notice that the route fan starts at the end of that wake, not at the original position.

This gives us the baseline: always take the newest forecast's fastest line. It also exposes the problem. An early choice can leave the boat badly placed when the next forecast arrives.

Add one affordable alternative

Now we can ask a more useful question. Instead of choosing the route that wins in one forecast, can we choose a nearby route that remains useful in several of them?

At each update, as many as four recent extended HRRR cycles provide candidate routes. We keep a decision only when at least three routes are complete. Each forecast's own optimum becomes one possible action. For every action, we calculate how much time it loses in the forecast or forecasts least favorable to it.

The rolling hedge chooses the least damaging route among those no more than 3 % slower in the newest forecast. If the newest forecast's fastest route is already the safest affordable choice, the hedge does nothing. Both navigators then sail for six hours through the same analysed weather and make their next decision from wherever they arrived.

These forecasts are successive operational runs, not independent ensemble members. Their disagreement shows how quickly the forecast answer changed. It does not describe every plausible atmosphere.

The historical limits matter too. A forecast hour enters the replay only after it was published. Older HRRR runs end at f18 or f36 rather than today's f48. If an adverse forecast is missing, we do not let that absence make a route look safer.

Did the hedge help?

We attempted the earliest fully covered racing evening in 22 edition-course cases. Sixteen cases from fourteen independent editions had every route, forecast and decision needed for a complete replay. If any comparison was missing, we did not score the case.

At the 3 % limit, the hedge finished about twenty seconds slower per case on average. The median difference was zero, and the 95 % interval ran from roughly three minutes slower to two minutes faster. We detected no average advantage.

The individual cases are less tidy. The best hedge gained seven minutes; the worst lost thirteen. The hedge chose a different route in ten of sixteen cases. Six cases finished unchanged because the deterministic route was also the safest affordable choice, or because the routes later converged.

So perhaps 3 % was the wrong price. We repeated the selection at four spending limits:

More freedom made the policy worse. A five-percent hedge lost 6.6 minutes per case on average; at ten percent, it lost 8.4. The route pays its insurance premium immediately, but the forecast difference may never arrive—or may disappear at the next update. On a short remaining passage, there may be no time to earn that cost back.

This tests one simple time-lagged HRRR policy, not ensembles in general. Every case uses one J/111 reference polar, the earliest evening with complete coverage, and the largest class with no more than forty boats. The cap excludes trophy and whole-fleet tracker groups, but it remains a design choice. The interval resamples editions, keeping paired Bayview courses together.

Forecast disagreement is not route choice

Why did uncertainty fail to produce value? Wind disagreement matters only when it sends the boat toward meaningfully different positions.

A ten-knot wind difference can leave every route in the same lane. There is nothing to hedge. A small direction change beside a shore or passage can split the routes, and that may matter much more.

This suggests two measurements the replay does not yet provide:

• route separation: how far apart do the forecasts place the boat?
• candidate capture: did any of those positions remain useful in the weather that followed?

The recent HRRR cycles can also move the same squall in the same wrong direction. In that case, the routes agree with one another and miss the useful position together. No selection rule can choose a future that is absent from every forecast.

Preserve reachable futures, not average routes

The answer is not to average the lines. Their average may cross a place that no forecast supports.

Compare complete routes instead. Pay for an alternative only when the forecasts lead to different positions, one position has a real downside, and the added time can still be recovered before the finish.

That returns us to the black wake. A navigator cannot revise the miles already sailed. A useful hedge protects the next decision without spending the rest of the race to get there.