Season overview

The 2026 Drum Corps International season is closed. ForeScorps forecast all 64 of its scored contests – 583 individual corps-results across World, Open, International and All-Age classes. The model called the winner of each event’s featured class 58 times out of 64: 90.6%. Its average error on a corps’ total was 1.396 points, with an RMSE of 2.493, and 88.7% of scores landed inside the 80% confidence range published before the show – against an 80% target. All figures on this page are scored contest-time: each event is measured against whichever forecast was live at that event’s contest time, under the model version active that day. No figure has been recomputed with hindsight. Per-event detail is available on the accuracy page; the models are documented in the methodology.

Model development over the season

Four model versions produced the 2026 forecasts. Published forecasts and baselines were never restated after the fact: each version went live on a date, took over a defined set of classes, and was scored from that date forward.

• Version 1 – effective June 27, 2026 in the model registry, launched publicly with the site on July 2. The original two-model ensemble: Model A, a statistical state-space model, and Model B, an AI/machine-learning model built on gradient-boosted trees, both trained on 13 seasons of DCI and DCA history.
• Version 2 – effective July 4, 2026. A hierarchical Bayesian model of the whole season, covering World Class and Open Class.
• Version 2.1 – effective July 12, 2026. In-season dynamics rebuilt so recent shows count for more; All-Age Open Class moved onto this model.
• Version 3 – effective July 19, 2026. Recency-calibrated: modern-era improvement curves, a validated first-show floor per corps, and the 80% bands tightened to match measured accuracy.
• Version 4 – effective August 1, 2026. The final-week calibrated ensemble: trailing bias correction, a season-scaled preseason anchor, a championship-week increment model, a step-model ensemble member, and recalibrated intervals.
• All-Age World Class and All-Age A Class stayed on Version 1 all season – the one documented case in which the newer model versions did not outperform Version 1, and the reason the original ensemble was still producing live forecasts in August.

The preseason baselines

Every version also carries a preseason baseline: that model run before the season began, with no 2026 scores in it. Those baselines answer a different question from the contest-time figures above – not how accurate each forecast was on the night, but how much of the season each model could have anticipated before it began. The table below is not a like-for-like comparison. Each row is scored on its own, differently-sized event set, so the rows are not comparable to one another and should not be read as one version beating another.

Two things in that table are worth noting. First, Version 4’s preseason baseline reproduces Version 3’s almost exactly by construction – every Version 4 correction is walk-forward, triggered by in-season observations, so at a preseason snapshot there is nothing for those corrections to act on. Its MAE of 3.401 against Version 3’s 3.335 is a difference in event set, not in model. Second, Version 1’s row is the most instructive number on this page: 89.1% of winners called from a standing start, with an MAE of 8.774 and band coverage of 34.1% against an 80% target. The original ensemble was largely correct about which corps would win while substantially overstating its certainty. Ranking proved an easier problem than score prediction, and calibrated confidence harder than either.

Model performance by version

Version 1: Model A, Model B and the ensemble

All three are shown on the same retrospective preseason footing – 64 events, 583 corps-results, no 2026 scores in the model.

The ensemble outperformed on the measure it was designed for: 89.1% of winners called, ahead of Model A’s 85.9% and Model B’s 84.4%. It did not outperform on point error: its MAE of 8.774 is fractionally higher than either component’s (8.768 and 8.423), and its RMSE of 9.797 falls between them. Blending the two models improved the predicted ordering of the field without making individual score predictions more accurate. All three shared the same 34.1% band coverage, as all three inherited the same overly narrow preseason intervals.

Versions 2 through 4 and the All-Age Version 1 track

These are the live, contest-time figures for the two model tracks that served forecasts during the season.

Versions 2 through 4, covering World Class and Open Class (with All-Age Open Class from July 12), produced the season’s headline accuracy: an average error of 1.146 points on 475 corps-results, roughly a seventh of the Version 1 preseason baseline’s, with 90.7% of winners called. The All-Age Version 1 track carried a higher error – 1.883 average error and an RMSE of 2.825 – and that is the result of forecasting a class with fewer shows and smaller panels. It is also the track with the best-calibrated bands on the site: 81.3% coverage against an 80% target is as close to the mark as any figure here. Version 1 continued to be used for All-Age World and A Class because it kept winning the walk-forward comparison in these classes.

Calibration: from 34.1% to 88.7%

The single largest improvement of the 2026 season was not in the forecasts themselves but in the uncertainty ranges around them.

An 80% confidence range carries a specific claim: approximately eight results in ten should land inside it. Version 1’s preseason bands captured 34.1% – a substantial calibration error, reflecting bands built for a world where the model believed it knew far more about June than it actually did. A forecast that identifies the correct winner but achieves only 34% coverage significantly overstates the confidence a reader should place in it.

Version 3’s recalibration corrected this: its own preseason baseline covers 84.7%, and Version 4’s 83.0%. Live, across the completed season, contest-time coverage finished at 88.7% – 88.8% for Versions 2–4 (World Class and Open Class) and 81.3% for the All-Age Version 1 track.

88.7% against an 80% target is not an exact win. The bands ended the season slightly too wide: the model was more conservative than the data required, at some cost in sharpness, since wider ranges produce weaker win probabilities. Of the two possible directions of calibration error, however, this is the preferable one: an interval that covers more than promised sacrifices precision, while an interval that covers 34.1% undermines confidence in the forecast itself.

Notable forecast successes

World Championship Finals: all twelve placements correct

At the DCI World Championship Finals on August 8, the forecast called all twelve World Class placements exactly, 1 through 12. Bluecoats were projected at 99.070 and won with 99.100 – a difference of 0.03 points on the season’s most prominent forecast. Event MAE across the full field was 0.462.

Additional results

• The champion was called from preseason. Bluecoats were the projected World Class winner before the season started, and they won it.
• Twenty-five consecutive correct calls to open the season. The featured-class winner was called correctly at 25 consecutive events, from the Barnum Festival on June 27 through July 14. The first miss came on July 16.
• Seventeen forecasts within 0.05 points. 17 of the season’s 583 corps forecasts landed within 0.05 points of the actual score.
• Lowest event-level error (MAE). Barnum Festival, June 27 – 0.434. World Championship Finals, August 8 – 0.462. DCI Houston, July 17 – 0.465. DCI Southern Mississippi, July 22 – 0.475. DCI Denton, July 16 – 0.536.
• Final class champions. World Class – Bluecoats, 99.100. Open Class – Blue Devils B, 81.725. All-Age World Class – Reading Buccaneers, season peak 96.200. Full standings are on the rankings page.

Forecast misses in detail

Six of the 64 featured-class winner calls were incorrect. Five occurred in contests forecast or decided within roughly one point, where the designation of a favorite carried limited meaning. The sixth was a clear model error. A recurring pattern of losses in near-even contests nonetheless indicates a gap in how the model identified and presented toss-up contests, and that is the principal finding of this section.

1. DCI Central Texas, July 16. Colts favored at 80.200 by a forecast margin of just 0.05 points; Blue Knights won with 82.500, beating Colts by 0.10 on the night. The season’s first miss, and it ended a 25-event streak by a tenth of a point.
2. Bushwackers Invitational, July 25. Hawthorne Caballeros favored at 92.000 on a 0.83-point forecast margin; Reading Buccaneers won at 91.500, by 1.20 on the night. The result ran counter to the model’s projected margin.
3. The Marion Open, July 30. The Battalion favored at 76.620; Gold won with 78.850, a margin of 0.45.
4. DCI Michigan, July 31. The reverse result the following night: Gold favored at 79.480; The Battalion won with 79.200, a margin of 0.10.
5. Open Class World Championship Prelims, August 3. Blue Devils B favored at 82.590, on the widest forecast margin of any missed call – 1.52 points. They finished third at 80.200, behind The Battalion at 80.800 and Gold at 80.400. This is the one miss that was not close: the favorite underperformed its own forecast by 2.4 points.
6. Open Class World Championship Finals, August 4. The favorite flipped to The Battalion at 81.150; Blue Devils B won with 81.725, by 0.225 against a 0.15-point forecast margin. The Prelims/Finals pair reversed on consecutive days; the Finals rematch, unlike the Prelims, was a genuine toss-up decided within the forecast margin.

The pattern is not spread evenly. The Battalion appears in four of the six misses – the Gold rivalry decided by 0.45 and 0.10 points on consecutive nights, then the Championships pair against Blue Devils B. The forecast margins are also informative: at DCI Central Texas the model favored Colts by 0.05 points, at DCI Michigan Gold by 0.31, and at the Open Class Finals The Battalion by 0.15 – toss-ups presented as favorites. The deficiency at those events was not forecast accuracy, as the projections were close, but presentation: a projected margin of a tenth of a point does not establish a favorite, and presenting it as one overstates the forecast’s certainty. The Prelims miss is the exception – a 1.52-point favorite finishing third is a model error rather than a narrow finish.

Partial hits

Winner accuracy on this page is measured on each event’s featured class. Below that headline, 10 multi-class events saw the model call some but not all of the class winners – those events show “Forecast hit – 50%” or “– 67%” on their event pages. At the other 52 scored events, every class winner was called correctly.

Largest score errors

Winner calls are the headline measure; individual score errors reveal the model’s specific failure modes.

• Golden Empire, Open Class, DCI Capital Classic, July 3 – forecast 24.868, actual 59.800, under by 34.9 points. The largest miss of the season by a wide margin, and a cold-start failure: a corps with minimal scoring history was anchored far too low by the early Version 1 preseason prior, with no corresponding widening of its uncertainty range. Version 4’s season-scaled preseason anchor targets exactly this failure mode.
• Mercedes Marching Band, International Class, Open Class Championship Finals, August 4 – forecast 75.170, actual 61.025, over by 14.1 points. The same problem from the opposite direction: a sparse-data international entrant, over-forecast rather than under.
• All-Age volatility. Sunrisers, All-Age World Class, at The Buccaneer Classic on July 18 – over by 11.3 points. Fusion Core, All-Age World Class, at DCI Birmingham on July 24 – under by 11.3. All-Age scores were the most volatile of the season, reflecting fewer performances to learn from and smaller judging panels.
• Worst events by MAE. DCI Capital Classic, July 3 – 6.955, driven by the Golden Empire result. Corps Encore, June 28 – 4.993. Drums Along the Rockies, June 27 – 3.968. All three fell in the season’s first week, before any 2026 results were available. The largest errors occurred where the model was working from historical data alone, consistent with the preseason coverage figure of 34.1%.

Potential changes for 2027

The following are candidate improvements identified from the 2026 season, each traceable to a specific figure above.

• Cold-start priors for sparse-history corps. The Golden Empire case: a 34.9-point miss resulted from a prior applied with high confidence to a corps with minimal scoring history. Sparse history should widen a forecast’s uncertainty range, not merely shift its center.
• Early-season uncertainty calibration. Preseason bands must be wide enough to achieve their stated coverage; Version 1’s covered 34.1% against an 80% target. Version 3 demonstrated that the correction is achievable, and future seasons should begin with calibrated bands rather than correcting mid-season.
• Toss-up presentation. When the forecast margin is inside roughly half a point, the contest should be presented as a toss-up rather than as a favorite and a challenger. Three of the six missed winner calls carried forecast margins of 0.31 points or less.
• Sparse-data methods for All-Age forecasting. The All-Age track carried the season’s highest average error at 1.883 and its two 11.3-point swings. The historical record is effectively complete with official scores from DCA through 2018. The constraint is not historical season data availability but the sparseness of scores within each season, as All-Age corps perform far fewer judged shows. The improvement path is forecasting methods designed for sparse data: stronger hierarchical pooling across corps and seasons, and priors better suited to short in-season score histories.
• Maintain the walk-forward version discipline. Versions 2 through 4 were introduced mid-season without restating any published forecast or baseline, which is what allows every figure on this page to be verified. The 2027 season will start from the Version 4 model, re-fit on 2026 results.

Figures on this page were generated from the site’s accuracy data, August 8, 2026, and are frozen – the 2026 season is closed and the data will not change. Per-event results are listed on the accuracy page, final standings on the rankings page, and the models are documented in the methodology. ForeScorps is an independent, unofficial project by B-Sharp AI Pte. Ltd. and is not affiliated with, endorsed by, or sponsored by Drum Corps International, Drum Corps Associates, or any drum corps.