fix: skip sources emptied by the eligibility filters in create_question_set

src/question_sampling/_prepare.py

@@ -165,6 +165,9 @@ def drop_questions_that_resolve_too_soon(
         is_na = dfq["forecast_horizons"] == "N/A"
         return dfq[~(empty_horizons | is_na)]
 
+    if dfq.empty:  # `apply(axis=1)` on an empty frame returns a Series that drops every column
+        return dfq
+
     forecast_due_date = forecast_datetime.replace(hour=23, minute=59, second=59, microsecond=999999)
     cutoff_date = forecast_due_date + timedelta(days=MIN_DAYS_TO_MARKET_RESOLUTION)
     resolves_too_soon = dfq.apply(
@@ -226,7 +229,7 @@ def drop_questions_that_resolve_too_late(
     Returns
         dfq (pd.DataFrame): Questions that resolve within the cap
     """
-    if not is_market_source(source):
+    if not is_market_source(source) or dfq.empty:
         return dfq
     cutoff_date = forecast_datetime + timedelta(days=MAX_DAYS_TO_MARKET_RESOLUTION)
     resolves_in_time = dfq.apply(

src/question_sampling/create_question_set.py

@@ -197,6 +197,9 @@ def create_question_set(
             forecast_datetime=forecast_datetime,
         )
         num_questions = len(dfq)
+        if dfq.empty:
+            logger.warning(f"No questions from {source_name} are eligible for sampling.")
+            continue
         QUESTIONS[source_name] = {
             "dfq": dfq,
             "num_questions_available": num_questions,

src/tests/unit/question_sampling/test_drop_questions_that_resolve_too_late.py

@@ -59,7 +59,7 @@ def test_empty_market_frame_stays_empty():
     result = drop_questions_that_resolve_too_late(
         source="polymarket", dfq=dfq, forecast_datetime=FORECAST_DATETIME
     )
-    assert result.empty
+    assert result.empty and list(result.columns) == list(dfq.columns)
 
 
 def test_data_questions_are_untouched():

src/tests/unit/question_sampling/test_drop_questions_that_resolve_too_soon.py

@@ -63,7 +63,15 @@ def test_empty_market_frame_stays_empty():
     result = drop_questions_that_resolve_too_soon(
         source="polymarket", dfq=dfq, forecast_datetime=FORECAST_DATETIME
     )
-    assert result.empty
+    assert result.empty and list(result.columns) == list(dfq.columns)
+
+
+def test_empty_data_frame_keeps_its_columns():
+    dfq = pd.DataFrame({"id": [], "forecast_horizons": []})
+    result = drop_questions_that_resolve_too_soon(
+        source="acled", dfq=dfq, forecast_datetime=FORECAST_DATETIME
+    )
+    assert result.empty and list(result.columns) == list(dfq.columns)
 
 
 def test_data_questions_without_horizons_are_dropped():

src/tests/unit/question_sampling/test_question_sampling.py

@@ -319,6 +319,16 @@ def resolution_dates(horizons: list[int]) -> list[str]:
         assert all(dates == resolution_dates([7, 30, 365]) for dates in by_source["fred"])
         assert all(dates == resolution_dates([30, 365]) for dates in by_source["acled"])
 
+    @pytest.mark.parametrize("emptied", ["polymarket", "fred"])
+    def test_source_whose_questions_are_all_filtered_out_is_skipped(self, bank, emptied):
+        questions, dfmeta = bank
+        questions[emptied] = questions[emptied].assign(resolved=True)
+
+        result = _create(questions, dfmeta)
+
+        sources = {q["source"] for q in result.llm_question_set["questions"]}
+        assert sources == {"polymarket", "kalshi", "fred", "acled"} - {emptied}
+
     def test_source_without_questions_is_skipped(self, bank):
         questions, dfmeta = bank
         questions["fred"] = pd.DataFrame(columns=constants.QUESTION_FILE_COLUMNS)