The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2545
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:51:37.178241Z
status: published
lens: g_space
sector: climate
headline: Station-specific thunderstorm forecasting using climate predictors with seasonal decomposition analysis in Bangladesh’s most vulnerable regions
dek: Thunderstorms pose significant threats to life, agriculture, and infrastructure in Bangladesh, particularly in the northeastern regions of Sylhet, Sreemangal, and Mymensingh. This study develops and rigorously evaluates localized thunderstorm frequency forecasting models for these high-risk stations using 40 years (1985-2024) of monthly data and thirty six modeling approaches encompassing classical time series (SARIMA/SARIMAX, ETS/ETSX), machine learning (ANN/ANNX, SVR/SVRX, XGBoost/XGBoostX), and deep learning…
gain_title: In a 40-year station-data study, adding five climate variables to seasonally adjusted machine learning models cut monthly thunderstorm forecast error (MASE) to 0.17-0.22 at Sylhet, Sreemangal and Mymensingh, over 80% below a univariate SARIMA benchmark at Sylhet.
problem_title: (none)
trace_subject: (none)
gain_reading: In a 40-year station-data study, adding five climate variables to seasonally adjusted machine learning models cut monthly thunderstorm forecast error (MASE) to 0.17-0.22 at Sylhet, Sreemangal and Mymensingh, over 80% below a univariate SARIMA benchmark at Sylhet.
gain_evidence: reflecting over an 80% reduction in MASE compared to the univariate SARIMA model (MASE = 0.7387)
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers compared 36 forecasting set-ups for monthly thunderstorm counts at three northeastern Bangladesh weather stations (Sylhet, Sreemangal and Mymensingh), using Bangladesh Meteorological Department records from 1985 to 2024. The best models paired seasonal decomposition (STL) with climate inputs. At Sylhet, STL-XGBoostX reached a MASE of 0.2214; at Sreemangal and Mymensingh, STL-SVRX reached 0.1715 and 0.1734.

Models were built on 1985-2016 data and tested on an unseen 2017-2024 period. Without climate inputs, the best model, STL-XGBoost, had a root-mean-square error of 2.53 at Sylhet, 2.09 at Sreemangal and 2.08 at Mymensingh, 41-45% lower than the same algorithm without seasonal adjustment. Adding temperature, humidity, cloud cover, rainfall and pressure lowered errors further for tree-based and kernel methods.

The work matters because lightning kills hundreds of people a year in Bangladesh, and the authors say station-level forecasting tools have lagged behind climatological description. The results point to April-August as the period when preparedness effort is most needed. The authors caution that forecast skill falls quickly with lead time and that their 2025-2028 projections illustrate model behaviour rather than reliable predictions.
limitation: The authors say forecast skill degrades rapidly with lead time, so the 2025-2028 projections illustrate model behaviour rather than reliable long-term predictions, and the projected rise at Sylhet conflicts with its historical decline.
tag: Evidence-backed gain
key_points: The dataset covers 480 months (1985-2024) of thunderstorm counts plus temperature, humidity, cloud cover, rainfall and pressure at three Bangladesh Meteorological Department stations. | Cloud cover (r = 0.83-0.88), atmospheric pressure (r = -0.79 to -0.85), temperature (r = 0.77-0.79) and rainfall (r = 0.74-0.78) correlated most strongly with thunderstorm frequency. | STL seasonal adjustment cut XGBoost RMSE from 4.1156 to 2.5297 at Sylhet, 3.32929 to 2.0943 at Sreemangal and 3.0182 to 2.0800 at Mymensingh. | LSTM deep learning performed poorly on raw data (MASE above 1.40); X-11 adjustment cut Sylhet LSTM MASE from 1.58 to 0.45. | Climate inputs did not always help: STL-LSTMX did worse than STL-LSTM at all three stations (Sylhet MASE 0.5409 vs 0.4922). | Sylhet showed a significant decline in thunderstorm frequency over 1985-2024 (p = 0.015); Sreemangal and Mymensingh showed no significant trend. | Peak activity fell in May and June, followed by April, July and August; Sreemangal's 2026 monsoon forecast was 100.5, the highest seasonal value.
rundown: Sylhet had the most thunderstorms over the period (6,029 events), ahead of Sreemangal (4,954) and Mymensingh (4,488). The authors link Sylhet's lead to its basin setting beside the Meghalaya and Tripura hills and moisture from the Bay of Bengal.

Forecasts were scored on 96 held-out months from 2017 to 2024, after tuning on 1985-2016 with expanding-window cross-validation at 1, 3, 6 and 12-month horizons. A raw XGBoostX model with climate inputs scored a Sylhet MASE of 0.3193 against 0.6839 without them. STL-SVRX narrowly beat STL-XGBoostX at Sreemangal and Mymensingh, whose MASE values were 0.2015 and 0.2040.

For 2025-2028, the tuned models projected near-zero winter and post-monsoon activity and a monsoon peak concentrated in May-September. Pre-monsoon forecasts for Sylhet rose from 47.6 in 2025 to 59.9 in 2027, and monsoon forecasts reached 90.3-94.9 at Sylhet and 67.4-72.8 at Mymensingh.

The authors say the study offers the Bangladesh Meteorological Department an operational framework that could be extended to more stations. Because the study had no long station-level socio-economic series, vulnerability is discussed only qualitatively, and satellite lightning data were not used.
sources:
- peer_reviewed | PLOS One | https://doi.org/10.1371/journal.pone.0358795 | 2026-09-24
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

9256b4f13b67b73968dd03b1ad2e2f56223d6b7160ab0f7f8c5b624ade1167db

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000