Improve error handling when no data available on forecast_date

[huisaddison]

I believe this is an evalcast issue. When we attempt to get_predictions() for a forecast_date during which there is no data available over the entire training window, covidcast returns a dataframe of 0 columns and 0 rows. evalcast then tries to data munge but raises a cryptic "issue not found in .data" error.

Suggestion:

• Catch the empty dataframe case and return a more informative message.

Backtrace:

arrange() failed at implicit mutate() step. 
* Problem with `mutate()` input `..1`.
✖ Column `issue` not found in `.data`
ℹ Input `..1` is `.data$issue`.
Backtrace:
     █
  1. ├─evalcast::get_predictions(...)
  2. │ └─`%>%`(...)
  3. ├─dplyr::bind_rows(.)
  4. │ └─rlang::list2(...)
  5. ├─purrr::map(...)
  6. │ └─evalcast:::.f(.x[[i]], ...)
  7. │   ├─base::do.call(...)
  8. │   └─(function (forecaster, name_of_forecaster, signals, forecast_date, ...
  9. │     └─modeltools:::forecaster(...)
 10. │       └─covidcast::aggregate_signals(df, dt = dt, format = "wide")
 11. │         └─covidcast:::apply_shifts(x, dt)
 12. │           └─base::mapply(apply_shifts_one, x, dt, SIMPLIFY = FALSE)
 13. │             └─(function (x, dt) ...
 14. │               └─covidcast::latest_issue(x)
 15. │                 └─covidcast:::first_or_last_issue(df, TRUE)
 16. │                   └─`%>%`(...)
 17. ├─dplyr::distinct(...)
 18. ├─covidcast:::issue_sort(.)
 19. │ ├─dplyr::arrange(df, dplyr::desc(.data$issue))
 20. │ └─dplyr:::arrange.data.frame(df, dplyr::desc(.data$issue))
 21. │   └─dplyr:::arrange_rows(.data, dots)
 22. │     ├─base::withCallingHandlers(...)
 23. │     ├─dplyr::transmute(new_data_frame(.data), !!!quosures)
 24. │     └─dplyr:::transmute.data.frame(new_data_frame(.data), !!!quosures)
 25. │       ├─dplyr::mutate(.data, ..., .keep = "none")
 26. │       └─dplyr:::mutate.data.frame(.data, ..., .keep = "none")
 27. │         └─dplyr:::mutate_cols(.data, ...)
 28. │           ├─base::withCallingHandlers(...)
 29. │           └─mask$eval_all_mutate(quo)
 30. ├─issue
 31. ├─rlang:::`$.rlang_data_pronoun`(.data, issue)
 32. │ └─rlang:::data_pronoun_get(x, nm)
 33. ├─rlang:::abort_data_pronoun(x)
 34. │ └─rlang::abort(msg, "rlang_error_data_pronoun_not_found")
 35. │   └─rlang:::signal_abort(cnd)
 36. │     └─base::signalCondition(cnd)
 37. ├─(function (e) ...
 38. │ └─rlang::abort(...)
 39. │   └─rlang:::signal_abort(cnd)
 40. │     └─base::signalCondition(cnd)
 41. └─(function (cnd) ...
> 

Reproducing example:

library(covidcast) # branch main
library(evalcast) # branch evalcast-killcards
library(modeltools) # branch main
library(dplyr)

## Setup 

# What are we forecasting?
response_source <- "jhu-csse"
response_signal <- "confirmed_7dav_incidence_prop"
incidence_period <- "day"
ahead <- 1:21
geo_type <- "hrr" 
forecast_dates <- seq(as.Date("2020-07-01"), as.Date("2021-01-31"), by = "day")

# Some quantgen parameters 
n <- 21               # Training set size (in days) 
lags <- c(0, 7, 14)   # Lags (in days) for features
no_pen_vars <- 1      # Variables to leave unpenalized (lastest response value)
nlambda <- 20         # Number of lambda values to consider in cross-validation
lp_solver <- "gurobi"
sort <- TRUE
nonneg <- TRUE

tau = c(0.05, 0.20, 0.50, 0.80, 0.95)

# Important: functions to considerably speed up data fetching steps. Only pull 
# recent data for each forecast date, depending on the training set size (and 
# other parameters for quantgen)
start_day_baseline <- function(forecast_date) {
  return(as.Date(forecast_date) - n - 4 + 1)
}

start_day_quantgen <- function(forecast_date) {
  return(as.Date(forecast_date) - max(ahead) - n - max(lags) + 1)
}

## Produce forecasts

# Quantile autoregression with 3 lags, or QAR3
t0 = Sys.time()
pred_quantgen1 <- get_predictions(
  forecaster = quantgen_forecaster, 
  name_of_forecaster = "QAR3 + CHNG_CLI3",
  signals = tibble::tibble(
                      data_source = c(response_source, "chng"),
                      signal = c(response_signal, "smoothed_adj_outpatient_cli"),
                      start_day = list(start_day_quantgen)),
  forecast_dates = forecast_dates, 
  incidence_period = incidence_period, 
  ahead = ahead, geo_type = geo_type, 
  tau=tau,
  signal_aggregation = "list",
  n = n, lags = lags, 
  nlambda = nlambda, no_pen_vars = no_pen_vars,
  debug = 'debug/QAR3+CHNG_CLI3',
  verbose=TRUE,
  lp_solver = lp_solver, sort = sort, nonneg = nonneg)
t1 = Sys.time()
print(t1-t0)