internal model check
Forecast validation
Did the predicted next 10 help us anticipate species that were new at a tracked source? This is the scorecard, not a count of moths or observations.
Network forecast evidence
Targets found shows how many predicted targets were first recorded during sheet-compatible hours in the following two weeks. Target coverage shows how many of those first-time records had been predicted. The repeated number is the overlap between those two groups. Results combine tracked sources and repeated 14-night tests; use the table below for Kingfisher Hollow alone.
Historical four-way backtest
Imagine writing a ranked ten-species prediction list on a past Monday, then checking what became newly recorded during sheet-compatible hours over the following 14 nights. We repeat that test across the network. Sheet-time is the production order; the other three show what changes when that filter or host evidence is removed. These are forecast results, not moth or observation totals.
Sheet-time production order
Keeps the seasonal order but moves a species behind neutral candidates when at least four pre-checkpoint timed records exist and half or fewer fall from 6 p.m. to 10 a.m. 72 windows with no species-level activity during sheet-compatible hours were not scored.
- targets found
- 17%
511 targets predicted across all stations; 86 found for the first time during sheet-compatible hours in the following two weeks
- target coverage
- 7%
1186 species first recorded at sheet-compatible hours in that period; 86 were predicted on the target list
- evidence
- 53 tests
340 active sheet-time nights across 14-night test windows
- typical useful rank
- #5
where a matching species sat in its original target list
Seasonal-only baseline
Ranks targets using only the frozen seasonal station-record evidence available at each checkpoint. 72 windows with no species-level activity during sheet-compatible hours were not scored.
- targets found
- 17%
511 targets predicted across all stations; 86 found for the first time during sheet-compatible hours in the following two weeks
- target coverage
- 7%
1186 species first recorded at sheet-compatible hours in that period; 86 were predicted on the target list
- evidence
- 53 tests
340 active sheet-time nights across 14-night test windows
- typical useful rank
- #5
where a matching species sat in its original target list
Host-only ranking
Uses the same seasonal candidate pool, but ranks it only by documented shared-host strength. Targets without a host match trail at the end. 72 windows with no species-level activity during sheet-compatible hours were not scored.
- targets found
- 7%
511 targets predicted across all stations; 35 found for the first time during sheet-compatible hours in the following two weeks
- target coverage
- 3%
1186 species first recorded at sheet-compatible hours in that period; 35 were predicted on the target list
- evidence
- 53 tests
340 active sheet-time nights across 14-night test windows
- typical useful rank
- #5
where a matching species sat in its original target list
Seasonal + host evidence
Uses the same checkpoint, target pool, and outcome window, with comprehensive host evidence used only to change ranking. 72 windows with no species-level activity during sheet-compatible hours were not scored.
- targets found
- 9%
511 targets predicted across all stations; 48 found for the first time during sheet-compatible hours in the following two weeks
- target coverage
- 4%
1186 species first recorded at sheet-compatible hours in that period; 48 were predicted on the target list
- evidence
- 53 tests
340 active sheet-time nights across 14-night test windows
- typical useful rank
- #4.5
where a matching species sat in its original target list
By station
Use this table for one station, such as Kingfisher Hollow. Stations with few moth nights have less evidence.
| Station | Historical sheet-time | Historical seasonal-only | Historical host-only | Historical seasonal + host | Historical sheet-time nights |
|---|
| Kingfisher Hollow |
26 / 140 (19%) |
26 / 140 (19%) |
11 / 140 (8%) |
16 / 140 (11%) |
112 |
| Bosque Neimi |
0 / 20 (0%) |
0 / 20 (0%) |
0 / 20 (0%) |
0 / 20 (0%) |
2 |
| Dombroskie Homestead |
4 / 121 (3%) |
4 / 121 (3%) |
3 / 121 (2%) |
5 / 121 (4%) |
99 |
| Monkey Run |
29 / 110 (26%) |
29 / 110 (26%) |
13 / 110 (12%) |
16 / 110 (15%) |
92 |
| Dove Dr |
20 / 70 (29%) |
20 / 70 (29%) |
6 / 70 (9%) |
8 / 70 (11%) |
25 |
| Second St |
not enough evidence |
not enough evidence |
not enough evidence |
not enough evidence |
0 |
| Durfee Hill |
not enough evidence |
not enough evidence |
not enough evidence |
not enough evidence |
0 |
| Ringwood Rd |
5 / 20 (25%) |
5 / 20 (25%) |
2 / 20 (10%) |
2 / 20 (10%) |
2 |
| Woodlawn |
2 / 30 (7%) |
2 / 30 (7%) |
0 / 30 (0%) |
1 / 30 (3%) |
8 |
Method notes
Historical four-way backtest: fourteen overlapping Monday checkpoints freeze the tracked-source records uploaded by local noon, then rebuild all four ten-species rankings from that same information. The next fourteen nights are the outcome period; only species new to that source and timestamped from 6 p.m. to 10 a.m. count as outcomes. Sheet-time is the production order: it preserves the seasonal ranking, but moves a species behind neutral candidates when at least four timed pre-checkpoint records exist and half or fewer were observed in that interval. Seasonal-only shows the unadjusted order. The host-only list uses host-association strength after the seasonal candidate pool is set; the combined list adds host evidence to the seasonal ranking. Exact shared host plants are weighted above broader genus overlap; no ranking can use observations uploaded after the checkpoint. Uploaded records do not consistently identify survey method, so sheet-compatible hours are a conservative timing proxy, not proof that an individual moth was physically present on a sheet.