fix: update flu+covid exclusions weights, minor pipeline tweaks

Author: dshemetov

R/forecasters/ensemble_linear_climate.R

@@ -128,14 +128,16 @@ make_ahead_weights <- function(aheads,
     tibble(forecast_family = "linear", ahead = sort(aheads), weight = 1 - ahead_weight_values)
   )
 }
+
 ensemble_weighted <- function(forecasts, other_weights) {
   forecasters <- unique(forecasts$forecaster)
-  full_weights <-
-    other_weights %>%
-    filter(
-      geo_value %in% unique(forecasts$geo_value),
-      forecaster %in% forecasters
-    ) %>%
+  filtered_weights <- other_weights %>%
+   filter(forecaster %in% forecasters) %>%
+   inner_join(
+      forecasts %>% distinct(forecaster, geo_value),
+      by = c("forecaster", "geo_value"),
+    )
+  full_weights <- filtered_weights %>%
     left_join(
       forecasts %>% mutate(ahead = target_end_date - forecast_date) %>% distinct(forecaster, ahead),
       by = "forecaster",

R/utils.R

@@ -178,8 +178,7 @@ data_substitutions <- function(dataset, disease, forecast_generation_date) {
     bind_rows(new_values)
 }
 
-parse_prod_weights <- function(filename = here::here("covid_geo_exclusions.csv"),
-                               forecast_date_int, forecaster_fn_names) {
+parse_prod_weights <- function(filename, forecast_date_int, forecaster_fn_names) {
   forecast_date_val <- as.Date(forecast_date_int)
   all_states <- c(
     unique(readr::read_csv("https://raw.githubusercontent.com/cmu-delphi/covidcast-indicators/refs/heads/main/_delphi_utils_python/delphi_utils/data/2020/state_pop.csv", show_col_types = FALSE)$state_id),

covid_geo_exclusions.csv

@@ -3,18 +3,27 @@ forecast_date,forecaster,geo_value,weight
 "2024-10-01", "all", "mp", 0
 "2024-10-01", "windowed_seasonal", "all", 3
 "2024-10-01", "windowed_seasonal_extra_sources", "all", 3
+"2024-10-01", "windowed_seasonal_extra_sources", "mo", 0
+"2024-10-01", "windowed_seasonal_extra_sources", "usa", 0
+"2024-10-01", "windowed_seasonal_extra_sources", "wy", 0
 "2024-10-01", "linear", "all", 3
 "2024-10-01", "linearlog", "all", 0
 "2024-10-01", "climate_base", "all", 2
 "2024-10-01", "climate_geo_agged", "all", 0.5
 "2024-10-01", "climate_quantile_extrapolated", "all", 0
 # feb 19
 "2025-02-19", "all", "mp", 0
-"2025-02-19", "windowed_seasonal", "all", 3
+"2025-02-19", "windowed_seasonal", "all", 0
 "2025-02-19", "windowed_seasonal_extra_sources", "all", 3
 # this is weighing the linear climate ensemble
 "2025-02-19", "linear_climate", "all", 0.0001
 "2025-02-19", "linear_climate", "ma", 3
+"2025-02-19", "linear_climate", "nh", 3
+"2025-02-19", "linear_climate", "nm", 3
+"2025-02-19", "windowed_seasonal", "pr", 3
+"2025-02-19", "windowed_seasonal", "mo", 3
+"2025-02-19", "windowed_seasonal", "wy", 3
+"2025-02-19", "windowed_seasonal", "usa", 3
 # creating linear climate
 "2025-02-19", "linear", "all", 3
 "2025-02-19", "linearlog", "all", 0

flu_data_substitutions.csv

@@ -8,3 +8,7 @@ nm, 2025-02-12, 2025-02-05, 200
 ca, 2025-02-12, 2025-02-05, 3893
 ms, 2025-02-12, 2025-02-05, 420
 ok, 2025-02-12, 2025-02-05, 850
+dc, 2025-02-19, 2025-02-15, 130
+ga, 2025-02-19, 2025-02-15, 700
+nm, 2025-02-19, 2025-02-15, 170
+ok, 2025-02-19, 2025-02-15, 600

flu_geo_exclusions.csv

@@ -2,11 +2,31 @@ forecast_date,forecaster,geo_value,weight
 # default values
 2024-10-01, all, mp, 0
 2024-10-01, windowed_seasonal, all, 3
+2024-10-01, windowed_seasonal_extra_sources, all, 3
 2024-10-01, linear, all, 0.5
 2024-10-01, linearlog, all, 0
 2024-10-01, climate_base, all, 0.5
 2024-10-01, climate_geo_agged, all, 0.25
 2024-10-01, climate_quantile_extrapolated, all, 0
+# feb 19
+2025-02-19, all, mp, 0
+2025-02-19, windowed_seasonal, all, 3
+2025-02-19, windowed_seasonal_extra_sources, all, 3
+2025-02-19, linear_climate, all, 3
+2025-02-19, linear, all, 0.5
+2025-02-19, linearlog, all, 0
+2025-02-19, climate_base, all, 0.5
+2025-02-19, climate_geo_agged, all, 0.25
+2025-02-19, climate_quantile_extrapolated, all, 0
+# state specific
+2025-02-19, linear, in, 3
+2025-02-19, linear, ks, 3
+2025-02-19, linear, ky, 3
+2025-02-19, linear, mt, 3
+2025-02-19, windowed_seasonal_extra_sources, mo, 0
+2025-02-19, windowed_seasonal_extra_sources, pr, 0
+2025-02-19, windowed_seasonal_extra_sources, us, 0
+2025-02-19, windowed_seasonal_extra_sources, wy, 0
 # feb 5
 2025-02-05, all, mp, 0
 2025-02-05, windowed_seasonal, all, 3

scripts/covid_hosp_prod.R

@@ -104,6 +104,14 @@ parameters_and_date_targets <- rlang::list2(
     score_report_rmd,
     command = "scripts/reports/score_report.Rmd"
   ),
+  tar_file(
+    covid_geo_exclusions,
+    command = "covid_geo_exclusions.csv"
+  ),
+  tar_file(
+    covid_data_substitutions,
+    command = "covid_data_substitutions.csv"
+  ),
   tar_change(
     nhsn_latest_data,
     change = get_socrata_updated_at("https://data.cdc.gov/api/views/mpgq-jmmr"),
@@ -163,6 +171,7 @@ forecast_targets <- tar_map(
             time_value = time_value - 3
           )
       } else {
+        covid_data_substitutions
         train_data <-
           nhsn_latest_data %>%
           data_substitutions(disease = "covid", as.Date(forecast_generation_date_int)) %>%
@@ -209,7 +218,7 @@ ensemble_targets <- tar_map(
   tar_target(
     name = geo_forecasters_weights,
     command = {
-      geo_forecasters_weights <- parse_prod_weights(here::here("covid_geo_exclusions.csv"), forecast_date_int, forecaster_fn_names_)
+      geo_forecasters_weights <- parse_prod_weights(covid_geo_exclusions, forecast_date_int, forecaster_fn_names_)
       if (nrow(geo_forecasters_weights %>% filter(forecast_date == as.Date(forecast_date_int))) == 0) {
         cli_abort("there are no weights for the forecast date {forecast_date}")
       }
@@ -269,9 +278,6 @@ ensemble_targets <- tar_map(
         ensemble_lin_clim %>% mutate(forecaster = "linear_climate"),
         ensemble_mixture_res %>% mutate(forecaster = "ensemble_mix"),
         ens_ar_only %>% mutate(forecaster = "ens_ar_only")
-        # TODO: Maybe later, match with flu_hosp_prod
-        # ensemble_mixture_res_2 %>% mutate(forecaster = "ensemble_mix_2"),
-        # combo_ensemble_mixture_res %>% mutate(forecaster = "combo_ensemble_mix")
       )
     }
   ),

scripts/flu_hosp_prod.R

@@ -14,15 +14,15 @@ end_date <- Sys.Date()
 # today, which is a Wednesday. Sometimes, if we're doing a delayed forecast,
 # it's a Thursday. It's used for stamping the data and for determining the
 # appropriate as_of when creating the forecast.
-# forecast_generation_dates <- Sys.Date()
+forecast_generation_dates <- Sys.Date()
 # Usually, the forecast_date is the same as the generation date, but you can
 # override this. It should be a Wednesday.
-# forecast_dates <- round_date(forecast_generation_dates, "weeks", week_start = 3)
+forecast_dates <- round_date(forecast_generation_dates, "weeks", week_start = 3)
 # If doing backfill, you can set the forecast_date to a sequence of dates.
-forecast_dates <- seq.Date(as.Date("2024-11-20"), Sys.Date(), by = 7L)
+# forecast_dates <- seq.Date(as.Date("2024-11-20"), Sys.Date(), by = 7L)
 # forecast_generation_date needs to follow suit, but it's more complicated
 # because sometimes we forecast on Thursday.
-forecast_generation_dates <- c(as.Date(c("2024-11-21", "2024-11-27", "2024-12-04", "2024-12-11", "2024-12-18", "2024-12-26", "2025-01-02")), seq.Date(as.Date("2025-01-08"), Sys.Date(), by = 7L))
+# forecast_generation_dates <- c(as.Date(c("2024-11-21", "2024-11-27", "2024-12-04", "2024-12-11", "2024-12-18", "2024-12-26", "2025-01-02")), seq.Date(as.Date("2025-01-08"), Sys.Date(), by = 7L))
 
 # Whether we're running in backtest mode.
 # If TRUE, we don't run the report notebook, which is (a) slow and (b) should be
@@ -125,6 +125,14 @@ parameters_and_date_targets <- rlang::list2(
     tar_file(
       score_report_rmd,
       command = "scripts/reports/score_report.Rmd"
+    ),
+    tar_file(
+      flu_geo_exclusions,
+      command = "flu_geo_exclusions.csv"
+    ),
+    tar_file(
+      flu_data_substitutions,
+      command = "flu_data_substitutions.csv"
     )
   ),
   make_historical_flu_data_targets(),
@@ -157,6 +165,7 @@ parameters_and_date_targets <- rlang::list2(
       } else {
         most_recent_result <- readr::read_csv("https://data.cdc.gov/resource/ua7e-t2fy.csv?$limit=20000&$select=weekendingdate,jurisdiction,totalconfc19newadm,totalconfflunewadm")
       }
+      flu_data_substitutions
       most_recent_result <-
         most_recent_result %>%
         process_nhsn_data() %>%
@@ -179,7 +188,7 @@ parameters_and_date_targets <- rlang::list2(
     name = nhsn_archive_data,
     change = get_s3_object_last_modified("forecasting-team-data", "archive_timestamped.parquet"),
     command = {
-      create_nhsn_data_archive(disease = "nhsn_covid")
+      create_nhsn_data_archive(disease = "nhsn_flu")
     }
   )
 )
@@ -260,7 +269,7 @@ ensemble_targets <- tar_map(
   tar_target(
     name = geo_forecasters_weights,
     command = {
-      geo_forecasters_weights <- parse_prod_weights(here::here("flu_geo_exclusions.csv"), forecast_date_int, forecaster_fn_names_)
+      geo_forecasters_weights <- parse_prod_weights(flu_geo_exclusions, forecast_date_int, forecaster_fn_names_)
       if (nrow(geo_forecasters_weights %>% filter(forecast_date == as.Date(forecast_date_int))) == 0) {
         cli_abort("there are no weights for the forecast date {forecast_date}")
       }
@@ -288,14 +297,13 @@ ensemble_targets <- tar_map(
     name = ens_climate_linear_window_season,
     command = {
       climate_linear %>%
+        mutate(forecaster = "climate_linear") %>%
         # Ensemble with windowed_seasonal and windowed_seasonal_extra_sources
         bind_rows(
           forecast_full_filtered %>%
             filter(forecaster %in% c("windowed_seasonal", "windowed_seasonal_extra_sources"))
         ) %>%
-        group_by(geo_value, forecast_date, target_end_date, quantile) %>%
-        summarize(value = mean(value, na.rm = TRUE), .groups = "drop") %>%
-        sort_by_quantile()
+        ensemble_weighted(geo_forecasters_weights)
     }
   ),
   tar_target(