barogram

Barogram is a pet forecast ensemble, a small collection of models I run for fun and to learn more about how forecasting actually works. Every three hours, they look at the latest readings from a backyard Tempest weather station in the Twin Cities, MN and a nearby NWS airport station, then each independently predict local temperature, dew point, pressure, and precipitation probability for the next 6 to 24 hours.

After each run, the previous predictions get scored against what actually happened. Models that beat a naive baseline earn more weight in the ensemble’s combined output; models that don’t are demoted toward a floor. The base models use simple approaches and none of them are impressive on their own. The ensemble is what makes them useful.

These forecasts are specific to that one station. This is a personal project running on data from my own equipment; it says nothing about conditions where you are.

Latest conditions in the Twin Cities

Ensemble Forecast — Generated at 17:09 on May 11, 2026

Verification

Overall Forecast Skill

Skill score vs. climatological mean, averaged across temperature, dewpoint, and pressure (plus Precip Prob BSS once enough rain events have been observed). 100% = perfect · 0% = matches climatological mean · negative = worse than climatological mean.

Skill Over Time

Daily forecast skill vs. climatological mean (0% line). Averaged across all variables. Default: ensemble, NWS, Tempest Forecast.

Forecast Skill by Lead Time

Skill score vs. climatological mean at each lead time for the selected variable. Negative = worse than climatology.

Brier Skill Score values are unreliable until enough precipitation events have been observed. With few or no rain events, the climo Brier reference approaches zero and scores become extreme.

Skill over time

Y-axis: MAE ÷ climo MAE per run. 1.0 = same error as climatological mean · below 1.0 = better · above 1.0 = worse. Grey: climo (long-dash) and persistence (dotted). Per-run detail: solid with rolling average overlay.

Weights

Skill-score weights computed by barogram tune. Each member is scored by how much it improves over a naive baseline; members that beat the baseline earn proportional weight, and those that don’t are floored or subfloored. Sector columns show how trust shifts across time-of-day.

Learnings

Tracked hypotheses that accumulate evidence over time. Thin data is expected early — the goal is to watch these relationships evolve.

Hypothesis A: Clearness persistence vs. pressure projection

Question: Does projecting the solar clearness index forward via pressure tendency (airmass_diurnal member 3) reduce temperature MAE compared to simply persisting it (member 1)? The weights table shows whether barogram tune tracks the better performer over time.

Status: +6h: m1 (clearness-only) leads by 0.8°F MAE; +12h: tied (10.5°F MAE). The ensemble is currently leaning on m1 (clearness-only) for temperature.

Hypothesis B: Solar clearness index vs. NWS sky cover

Question: Does the Tempest station’s solar-derived clearness index (k) agree with NWS-reported sky cover? Each point is a daily average. What to look for: the two lines should move inversely (clearness drops on cloudy days, sky cover rises). If they move together or persistently diverge, there may be a sensor issue or a real local microclimate difference between the Tempest site and KMSP. NWS sky cover is never used as a model input — this is validation only.

Status: clearness and sky cover are moving in opposite directions as expected — cloudier days show lower clearness. Relationship is consistent (31 days of data).

Hypothesis C: Is the ensemble closing the gap on its best member?

Question: The ensemble is currently worse than climo_deviation on temperature at every lead. The line shows the rolling gap (ensemble MAE − climo_deviation MAE, 10-run mean) over time. What to look for: the line trending toward or below zero — that means the ensemble is learning to match or beat its best member. A flat or rising line means the weighting is not converging.

Status: +6h gap +2.2°F (↑ diverging); +24h gap -1.7°F (↓ converging). Ensemble still trails climo_deviation.

Hypothesis D: pressure_tendency — best and worst simultaneously

Question: pressure_tendency is the best model for dewpoint at all leads, but its pressure MAE climbs steeply (40+ hPa at 24h vs persistence’s 5 hPa). Both lines are shown at +12h with a 10-run rolling mean. What to look for: the two lines diverging — low dewpoint, high pressure. That’s expected and confirms the model design trade-off. If pressure MAE starts dropping back toward dewpoint level, something has changed.

Status at +12h: dewpoint MAE 4.6°F, pressure MAE 4.4 hPa. Pressure error is surprisingly low. If it stays this close to normal, the model's pressure sacrifice may have shrunk.

Hypothesis E: How long does the climo_deviation advantage last?

At +6h, climo_deviation beats persistence by ~1.9°F; by +24h that gap has shrunk to ~0.5°F. Lines are 10-run rolling means. What to look for: the two lines converging at +24h (gap approaching zero) while staying well separated at +6h. If they converge at +6h too, the recency signal has lost value. A seasonal shift (gap changes in summer vs winter) would also be meaningful.

Status: +6h: climo 5.8°F vs persistence 13.3°F (gap +7.5°F); +24h: climo 6.9°F vs persistence 8.7°F (gap +1.8°F). Advantage decaying with lead as expected.

Hypothesis F: Model specialization map

Which base model wins each (variable × lead) cell? Hover for MAE and sample size. What to look for: does the ensemble weighting actually reflect this map? If the ensemble underperforms for a variable, check whether the dominant model here gets high weight in that column.

Status: full_state_analog has the lowest error in 4 of 16 variable/lead combinations. The ensemble is not giving extra weight to the model that's actually winning (5 of 16 combinations agree, 31%).

Hypothesis G: Does diurnal_curve ever beat climo_deviation?

diurnal_curve models the daily temperature cycle explicitly. climo_deviation wins at every lead right now by anchoring to recent deviations from climatology. Lines are 10-run rolling means. What to look for: diurnal_curve closing the gap, especially at overnight leads (+18h/+24h) where solar effects matter less. If it never closes, the recency signal in climo_deviation is the explanation — not the diurnal cycle.

Status: +6h: diurnal 7.0°F vs climo_dev 5.8°F (gap +1.2°F); +24h: diurnal 4.4°F vs climo_dev 6.9°F (gap -2.5°F). diurnal_curve closes gap and leads at longer horizons.

Hypothesis H: Does trend window length have an optimal size?

Question: For each lead time, is there an optimal trend window where MAE is minimized — short enough to capture the recent signal, long enough to avoid noise? Or does skill simply improve monotonically with more history? multivariate_trend members span 1–48h windows; each point here is the all-time avg MAE for one (member, lead) pair. Note: early data for short-window members at long leads reflects pre-fix era forecast errors — those members are now restricted to appropriate leads.

Status: +6h best: 3h window (14.9°F MAE); +12h best: 1h window (30.9°F MAE); +18h best: 6h window (16.1°F MAE).

Latest Forecast Run

Observation History

Tempest Weather Station

NWS: Minneapolis-St. Paul Intl (KMSP)