CastScience

Science Info and Calculations

Validation

How the engine is checked: against reference calculations, against tournament results, and against measurements on the lake.

In plain words

Three kinds of checks, each answering a different question:

  • Golden tests: does the app compute what the reference scripts compute? Every formula is ported from Python scripts, and the ports must reproduce their results.
  • The tournament backtest: does the plan point to the water where tournaments were actually won, using only what was known before each event?
  • Measurements: do the CastScience Wave Prediction, over-water wind and lake temperature match what the mid-lake buoy and the satellite lake temperature measure?

Golden tests

Run now, in this browser, on the same engine code the plan uses, against the reference values in data/golden/engine_golden.json:

CheckLargest differenceToleranceResult
Wave growth, 45 cases: SPM height0.0005 ft0.05 ftpass
Wave growth: SPM period0.0005 s0.05 spass
Wave growth: Young-Verhagen height0.0005 ft0.05 ftpass
Water temperature projection, 12 days, trip analysis coefficients0.0005 F0.01 Fpass
Water temperature projection, 12 days, monthly coefficients (lake record)0.0005 F0.01 Fpass
Largest wave ratio, sqrt(ln 1000 / 2)1.8585 shown as 1.86roundingpass

The full test suite (npm test, before every change) adds more: the area wave table (384 rows) and the waves on the boat ride from the Metropark (288 rows) of the trip analysis to 0.1 ft; fetch for 12 areas in 8 directions to 0.1 km; the Oct 7-11 regression and climatology; the daily front rule's 15 front days on Selfridge data; sunrise and sunset; season phases and depth bands; area scoring, lure scoring and the day schedule on Lake St. Clair scenarios; and that the figures in the wave disclaimer follow from the wave model.

Source: packages/engine/test.

Tournament backtest

For each in-lake tournament with a known winning area, the engine builds the plan for day 1 from the weather that actually happened (hourly observations at Selfridge, gaps filled from Detroit Metro, land wind times the lake record's land-to-water factor for the month, 0.95 in June to 1.37 in October, for over-water wind), the GLSEA lake temperature, and a generic angler profile: deep-V, 2.0 ft wave limit, both licences, all lures, the Metropark launch. Only tournaments before each event count as evidence (no peeking). The gate passes when the winner's area is in the plan's top 3 and the winner's main lure family is in the top 3 lures, each for at least 8 events.

Latest run, 2026-11-18: PASS. Winner's area in the top 3: 8 of 11 (needed 8). Main lure family in the top 3: 8 of 10 events that reported their baits (needed 8).

  • For comparison, a plan that simply picks the 3 areas with the most earlier wins hits 6 of 11.
  • With the app's crossing rule on, as for a user: 9 of 11. The gate scores with it off, because tournament anglers crossed open water in winds the rule does not allow.
Event, day 1LakePhase, frontWinner's areaPlan's top 3 (score)HitWinner's lureTop 3 lures thereHit
Bassmaster Elite 2015, 2015-08-2769.3 FSummer, stableCA1US5 67.1, US1B 69.2, CA4 66.6nodrop shotPaddle-tail swimbait, Drop shot, Tubeyes
Bassmaster Elite 2017, 2017-08-2471.9 FSummer, second day after a frontCA2US1B 69.2, US5 64.9, CA4 63.9notubeDrop shot, Tube, Paddle-tail swimbaityes
FLW Tour 2018, 2018-06-2872.5 FSummer, day after a frontCA1CA4 71.9, CA1 74.4, US1A 70.7yestubeDrop shot, Tube, Paddle-tail swimbaityes
BFL Michigan Division 2018, 2018-07-1475.3 FSummer, stableCA1CA1 81.9, US1A 70.6, US5 67.5yesnot reportedTube, Drop shot, Paddle-tail swimbaitn/a
BFL Michigan Division 2019, 2019-08-0374.7 FSummer, front passingCA4CA1 85.7, US1A 70.7, CA4 71.9yestubeDrop shot, Tube, Paddle-tail swimbaityes
Bassmaster AOY Championship 2019, 2019-09-2965.6 FEarly fall, second day after a frontCA2US1B 79.2, US5 62.8, CA2 61.5yescrankbaitTube, Drop shot, Suspending jerkbait 110-115 mmno
Bassmaster Elite 2020, 2020-08-2074.4 FSummer, before a frontUS1A or US1BUS1B 74.2, US3 63.5, US5 65.9yesdrop shotDrop shot, Paddle-tail swimbait, Medium-diving crankbaityes
MLF Bass Pro Tour Stage Six 2023, 2023-06-2968.6 FPost-spawn, stableCA1US2 74.3, US4B 67.8, CA5 68.2nodrop shotDrop shot, Tube, Paddle-tail swimbaityes
Bassmaster Elite 2023, 2023-07-2775.2 FSummer, stableCA4CA1 85.7, CA4 79.3, US1A 74.3yesdrop shotDrop shot, Tube, Paddle-tail swimbaityes
Bassmaster Open 2024, 2024-07-1173.6 FSummer, day after a frontUS1A or US1B or CA4CA4 82.9, CA1 85.2, US1A 74.2yesdrop shotDrop shot, Tube, Paddle-tail swimbaityes
Bassmaster Elite 2025, 2025-08-0775.5 FSummer, stableUS1A or US1BUS1A 81.9, CA4 83.8, CA1 82.2yesfinesseDrop shot, Tube, Paddle-tail swimbaitno

Limits of this test: the observed weather stands in for the forecast, so it tests the scoring, not forecast error. The record is mostly June-August events; only one event falls in early fall. Scoring was tuned in one recorded round and then frozen (2026-10-01), because further tuning on the same 11 events would make the gate meaningless; the next check is on new data: the 2026 events and anglers' logbooks.

Source: Full report: reports/backtest-20261118.md (npm run backtest, scripts/backtest/run.ts).

Checks against measurements

Calibration run of 2026-10-03. Buoy 45147 sits in mid-lake (36 directions of open water, 10 to 27 km; depth 6 m) and reports wind, waves and air and water temperature every hour of its season. It is paired with Selfridge (KMTC), the NWS Lake St. Clair forecast and the GLSEA lake temperature.

Wind over water

Over-water wind divided by land wind, hours with a land wind of 10 mph or more, 2000-2026 (22,102 hours in 23 seasons): 1.19 for all seasons, 95% interval 1.16-1.22. The app now uses one factor per month (Wind over water); before, it used 1.2 in every month.

MonthFactor95% intervalHours
Apr1.171.10-1.24624
May1.040.97-1.101,928
Jun0.950.92-0.992,943
Jul0.980.96-1.012,966
Aug1.101.06-1.143,110
Sep1.271.22-1.313,065
Oct1.371.32-1.413,936
Nov1.411.37-1.443,022
Dec1.541.38-1.63508

Over-water wind predicted from land wind of 10 mph or more: root-mean-square error 5.5 mph with the monthly factors, 6.0 mph with 1.2. In October the single factor ran 2.2 mph low on average, in July 2.7 mph high.

Waves

Modeled significant wave height against the buoy's, 83,507 hours in 2003-2026, at 4.5 m depth as the app uses for open water. Hours with 1 ft or more measured or modeled; bias = model minus measured:

Model and windHoursMeasured, meanModel, meanBiasRMS errorMeasured 25% or more above the modelSame, 2 ft or more
SPM (the app's model), buoy wind37,7951.46 ft1.51 ft+0.04 ft0.49 ft10.8%17.3%
Young and Verhagen, buoy wind34,4751.52 ft1.36 ft-0.16 ft0.53 ft25.1%58.5%
SPM, Selfridge wind times the monthly factor22,8141.44 ft1.22 ft-0.22 ft0.72 ft45.9%42.3%
SPM, Selfridge wind times 1.2 (before the 2026-10-02 calibration)24,5241.39 ft1.21 ft-0.18 ft0.74 ft44.7%55.5%

With the wind measured on the water, the app's model is close on average; most of the error in a plan comes from the wind. The NWS forecast for the first 12-hour period (14,533 periods, 2003-2026), against the highest wave the buoy measured in the period:

ForecastBiasHighest wave above the forecast25% or more aboveSame, periods with waves above 2 ftPeriods above 2 ft the forecast put at 2 ft or less
NWS wave forecast, upper value+0.62 ft14.1%12.2%21.5%32.0%
The app's model at the NWS forecast wind+0.07 ft39.2%19.2%38.8%25.9%

These figures are the basis of the wave disclaimer. Measured at one point: waves near shore and at the long-fetch southeast shore are not checked.

Lake temperature

The projection from every ice-free day of 1995-2026 with the weather that happened, against GLSEA (F; monthly coefficients now, October-only before the 2026-10-02 calibration):

Days aheadCasesRMS error beforeRMS error nowBias now80% of errors between
16,4920.59 F0.60 F-0.07 F-0.7 to +0.7 F
36,2911.22 F1.18 F-0.20 F-1.6 to +1.3 F
75,9721.81 F1.63 F-0.43 F-2.4 to +1.6 F
145,5312.25 F1.89 F-0.68 F-2.9 to +1.6 F

Fitted on even years and tested on odd years, and the reverse, 7 days ahead: 1.67 F against 1.81 F. The Sept 18-29, 2026 northeast blow, hindcast: at most 1.8 F warm (was 2.7 F). With forecast weather the errors are larger; the band the app shows comes from 31 years of GLSEA alone (Water temperature).

Source: reports/calibration-1.8-20261002.md; reports/calibration-latest.json (npm run calibration, scripts/calibration). Buoy 45147: Environment and Climate Change Canada; archive Fisheries and Oceans Canada MEDS via CIOOS Pacific; contains information licensed under the Open Government Licence - Canada. Selfridge, Detroit Metro and the archived NWS forecasts: Iowa Environmental Mesonet. GLSEA: NOAA GLERL.

Plan confidence

Each plan carries a confidence label for how far its data can be trusted, not the chance of a catch:

  • High: the date is 3 days ahead or less, the lake temperature is inside the tournament record's 65.7-76.3 F, and the top area has 2 or more supporting records.
  • Low: 8 or more days ahead (past the 7-day forecast, climatology only), or the lake more than 5 F outside that range, or a key data feed out of date.
  • Medium: everything else.

Source: Engine: planConfidence in packages/engine/src/confidence.ts.

Known limits

  • Area positions are approximate centers or outlines from published descriptions, not waypoints.
  • The water temperature coefficients are fitted by month (calibration of 2026-10-02) on the weather that happened; a plan projects with forecast weather, whose errors add to the model's.
  • The land-to-water wind factor is fitted by month from buoy 45147 against Selfridge (calibration of 2026-10-02); a single year's factor can differ from the average by about 0.1.
  • Waves are checked at one point, the mid-lake buoy; gusts, wind shifts, wakes, reflection from walls and current against wind are not modeled.
  • Season-phase limits, front thresholds, lure attributes and scoring weights are starting values; the backtest and logbooks check them.
  • The NWS forecast has no pressure, so forecast days use the daily front rule.
  • The clarity model is not yet calibrated, and river plumes are a constant until river gauges are added.
  • The tournament record is mostly June-August; one event falls in early fall.