Sitrep: ensemble moisture forecast for the fire seasons ahead: Bogota Cerros Orientales and the southern Amazon

Sitrep: ensemble moisture forecast for the fire seasons ahead

Bogota Cerros Orientales and the southern Amazon · issued at week 29 of 2026, 27 July 2026 · next reissue at week 30

Bogota, elevated Amazon, watch

Bogota. The model issued today gives a 71 per cent chance that the December 2026 to March 2027 season falls into the driest third on record, against a climatological rate of 33 per cent, and a 74 per cent chance of at least three weeks below one sigma. The Cerros index sits at -0.36 and the path ensemble puts a 50 to 53 per cent weekly chance of being at or below the fire response break of -0.56 through August and September.

Amazon. The southern Amazon read +1.16 in week 29, a wet week, which is why the weekly chance of sitting below the break stays near zero into early September. That single week sits on top of a dry season to date: the six site mean over weeks 18 to 29 is -0.32 and the year to date is -0.35, and weeks 22 and 23 were near -1.2. The seasonal model gives 51 per cent for the 2026 dry season falling in the driest third, above the 33 per cent climatological rate, and the land is the reason rather than the ocean. Seven of the fourteen network sites are below -0.5 sigma against a normal of between two and three, which is the 92nd percentile of the record, and within the southern Amazon the Cerrado, Xingu, arc of deforestation and Amazon south sites are all dry. Decomposing the August issue into one term per predictor puts 87 per cent of the movement in the soil moisture network and 13 per cent in the Pacific.

The Pacific is worth stating separately, because it is doing something dramatic and it is not what is driving this number. The Nino 3.4 index has moved from -0.37 in January to +0.98 in May 2026, a February to May rise of +1.12 that is the largest for those months in the record since 1950, just ahead of 1997. That rise is the substance of the Bogota warning above, where the Pacific carries 64 per cent of the movement. It is not the substance of the Amazon one. Over 76 seasons the Pacific alone discriminates Amazon dry seasons at 0.49 and 0.41 area under the curve, at and below a coin, while antecedent soil moisture alone reaches 0.76 to 0.87.

What this is for. Early warning as Bogota and the Amazon begin to lose soil moisture. The value of the lead time is that prevention is cheap and repair is not, so a warning issued in July that is only moderately confident is still worth acting on if the action is cheap and reversible.

Where both basins stand today

Everything below uses data up to and including week 29 of 2026 and nothing after it. The Bogota targets refer to the season that opens in December 2026 and closes in March 2027. The Amazon targets refer to the 2026 dry season, weeks 35 to 48.

target	forecast index	critical value	probability	AUC	Brier	configuration
Bogota season mean	-0.41	-0.18	71%	0.800	0.166	ridge
Bogota dry weeks	+4.89	+3.00	74%	0.788	0.175	ridge
Amazon season mean	-0.28	-0.15	51%	0.787	0.167	ensemble
Amazon core	-0.34	-0.20	53%	0.770	0.169	ensemble

AUC and Brier are the out of sample scores of that configuration over 76 seasons, nested leave one season out. Configuration says which of the two candidates won that target: ridge is the penalised regression on the mode extended predictor pool, ensemble adds the analog member and the direct logistic through a guarded two level stack.

The warning as it has evolved this season

The forecast is rebuilt from scratch at every week, with the predictor windows and the fourteen site modes truncated at that week, so the line below is what would have been issued each Monday rather than a smoothed version of today's answer.

Hover any point for its value. The dashed line is the climatological rate: a forecast on that line carries no information. The Bogota probability was highest in May, when only two weeks of the window had been observed and the estimate was least reliable. It has settled near 70 per cent as the window filled.

How much the forecast can be trusted at each week

Discrimination rises through the window, which is the honest reason not to over read the May issue. Bogota is usable from about week 22. The Amazon is at chance until week 21 and only becomes useful after week 27, so an Amazon warning issued in May is not yet worth acting on.

When, not only whether

The path ensemble takes every historical season's trajectory after week 29, adds it to this year's observed level, and weights each season by how close its outcome was to what the seasonal model forecasts for 2026. The result is a distribution over the week the index first reaches the fire response break at -0.56.

Bogota, 36 weeks ahead, to the end of March 2027.

Amazon, 20 weeks ahead, to the end of the dry season.

basin	index now	chance of reaching the break	median crossing	eighty per cent interval
Bogota Cerros	-0.36	96%	week 31, around 30 July 2026	week 30 to week 34, 23 July to 20 August
southern Amazon	+1.16	14%	week 38, around 17 September 2026	week 34 to week 45, 20 August to 5 November

Crossing dates are conditional on a crossing happening at all. The Bogota number is high because the Cerros index crosses this level in most seasons at some point over a 36 week horizon, so the informative quantity there is not whether but how deep and how long, which is the next table.

How bad a dry spell could get

A dry spell is a run at or below one sigma, separated from the next by at least four weeks. The generalised Pareto fit on spell peaks answers a different question from the seasonal model: given a spell has started, how far does it go.

basin	depth now	spell underway	chance of reaching 2 sigma	chance of reaching 2.5 sigma	tail shape
Bogota Cerros	+0.36	no, 89% chance one starts	17.4%	3.3%	bounded, floor near 3 sigma
southern Amazon	-1.16	no, 10% chance one starts	0.7%	0.2%	effectively exponential, no ceiling

The two basins differ in kind. Bogota's tail is bounded, so a spell that starts has a limited worst case and the operational question is duration. The Amazon tail has no comparable ceiling, so the same nominal warning carries a longer right hand risk there, and that is the case for acting earlier in the Amazon even when the probability is lower.

What got worse, and where the ensemble is not used

The ensemble is not adopted everywhere. Two of the eight operational models score worse than the single model they were meant to replace, and on those the old model stays in service. Reporting this is part of the product.

model	ensemble AUC	baseline AUC	ensemble Brier	baseline Brier	decision
bog_dry august	0.788	0.799	0.175	0.184	single model retained
amz_mean september	0.885	0.906	0.139	0.124	single model retained

Three further things got worse along the way and were fixed rather than hidden. Stacking three members without a guard scored below the best single member on five Bogota targets, because the weights were fitted on 75 inner points. Adding the one standard error guard recovered most of that but pushed bog_dry august down to 0.721 AUC, because the adoption rule was only comparing against the frozen baseline and had no way to keep a configuration that had already beaten both. And calibrating with an intercept and a slope removed the Amazon over forecast but cost AUC on six of eight targets, because a slope refitted in every fold stops being a single monotone map and reorders the seasons. The final build keeps two candidates per target, picks on AUC, and calibrates with an intercept only.

Full scorecard

model	config	skill	baseline skill	AUC	baseline AUC	Brier	baseline Brier	decision
bog_mean august	ridge	+0.227	+0.108	0.800	0.760	0.166	0.185	adopted
bog_mean september	ridge	+0.258	+0.260	0.823	0.784	0.162	0.174	adopted
bog_mean december	ridge	+0.382	+0.352	0.890	0.885	0.122	0.134	adopted
bog_dry august	ridge	+0.155	+0.127	0.788	0.799	0.175	0.184	not adopted
bog_dry december	ridge	+0.233	+0.250	0.855	0.825	0.150	0.168	adopted
amz_mean august	ensemble	+0.247	+0.084	0.787	0.691	0.167	0.207	adopted
amz_mean september	ridge	+0.339	+0.358	0.885	0.906	0.139	0.124	not adopted
amz_core august	ensemble	+0.203	+0.086	0.770	0.730	0.169	0.192	adopted

Skill is one minus the ratio of forecast error to climatology error on the continuous index. All eight are nested leave one season out over 76 seasons, 1950 to 2025, with penalty, neighbour count, stacking weights and calibration all chosen inside the fold.

The weekly table behind the charts

issue week	Bogota mean p	Bogota dry p	Amazon mean p	Amazon core p	Bogota mean AUC	Bogota dry AUC	Amazon mean AUC	Amazon core AUC
18	85%	95%	57%	50%	0.715	0.754	0.497	0.534
19	96%	97%	54%	52%	0.711	0.751	0.499	0.537
20	98%	99%	63%	62%	0.729	0.774	0.478	0.512
21	97%	98%	67%	53%	0.736	0.748	0.560	0.525
22	90%	90%	72%	70%	0.749	0.776	0.639	0.666
23	88%	88%	75%	73%	0.748	0.782	0.648	0.659
24	86%	81%	71%	69%	0.742	0.783	0.517	0.590
25	80%	78%	52%	56%	0.752	0.781	0.620	0.686
26	78%	76%	42%	43%	0.762	0.780	0.615	0.697
27	82%	77%	44%	58%	0.769	0.779	0.692	0.718
28	73%	76%	43%	44%	0.793	0.786	0.746	0.763
29	71%	74%	51%	53%	0.800	0.788	0.787	0.770

What to watch in the next four weeks

Three things decide whether this warning firms up or relaxes. First, the Cerros weekly index: two consecutive weeks at or below -0.56 in August move Bogota from watch to alert, and the path ensemble says that is more likely than not. Second, and for Bogota only, the Pacific: if the Nino 3.4 index holds above +0.5 through the June to August season the Bogota probability firms up with it, because the Pacific carries most of the Bogota signal at this lead. For the Amazon the thing to watch is not the Pacific but the southern Amazon sites themselves, which is where 87 per cent of that forecast comes from: a return to the weeks 22 and 23 level near -1.2, sustained rather than for one week, is what would move the Amazon number, and the wet week just observed is what is currently holding it down. Third, the September issue for the Amazon, which uses the observed August window and historically scores 0.885 AUC against 0.787 today, so the single most valuable thing that happens between now and then is simply the passage of five more weeks of observation.

The next scheduled reissue is week 30. Nothing in this sitrep requires new data collection: every input is a public product already in the pipeline.

Early warning system for wildfire severity, Bogota Cerros Orientales and the southern Amazon. Tecnologico de Monterrey. Contact alejandrofeged@tec.mx. ORCID 0000-0001-8649-4213. Generated 27 July 2026 from forecast_ensemble.py. All scores nested leave one season out, n = 76 seasons.