diff --git a/src/gluonts/__init__.py b/src/gluonts/__init__.py
index 30cbec26..d863292c 100644
--- a/src/gluonts/__init__.py
+++ b/src/gluonts/__init__.py
@@ -16,6 +16,7 @@
 
 from pkgutil import extend_path
 
+from . import _pandas_compat  # noqa: F401
 from .meta._version import __version__
 
 __all__ = ["__version__", "__path__"]
diff --git a/src/gluonts/_pandas_compat.py b/src/gluonts/_pandas_compat.py
new file mode 100644
index 00000000..1cf209ea
--- /dev/null
+++ b/src/gluonts/_pandas_compat.py
@@ -0,0 +1,138 @@
+"""
+Compatibility shim for pandas 3.0+.
+
+pandas 3.0 removed deprecated frequency aliases (H, T, S, M, A, Q, etc.).
+gluonts 0.17.0 was written for pandas <3 and uses these old aliases
+throughout its source code and test suite. Upstream caps pandas at <3
+(see https://github.com/awslabs/gluonts/pull/3321), but nixpkgs ships
+pandas 3.0+ only.
+
+This module monkey-patches to_offset, pd.Period, pd.period_range,
+pd.date_range, and DatetimeIndex.to_period to translate old-style
+frequency strings to their new-style equivalents before they reach
+pandas. The mappings differ between offsets (which use ME/QE/YE for
+end frequencies) and periods (which use M/Q/Y).
+
+It also restores the Tick.delta property removed in pandas 3.0.
+"""
+
+import re
+
+import pandas as pd
+import pandas.tseries.frequencies as _freq_module
+from pandas.tseries.offsets import Tick as _Tick
+
+# Old to new mappings for to_offset and date_range.
+# End frequencies get an E suffix; AS becomes YS.
+_OFFSET_MAP = {
+    "H": "h",
+    "T": "min",
+    "S": "s",
+    "M": "ME",
+    "m": "ME",
+    "A": "YE",
+    "Y": "YE",
+    "Q": "QE",
+    "AS": "YS",
+}
+
+# Old to new mappings for pd.Period and pd.period_range.
+# Period uses short names for end frequencies (M, Q, Y) and does not
+# support start variants (MS, QS, YS, AS) at all.
+_PERIOD_MAP = {
+    "H": "h",
+    "T": "min",
+    "S": "s",
+    "M": "M",
+    "m": "M",
+    "A": "Y",
+    "AS": "Y",
+    "YS": "Y",
+    "MS": "M",
+    "QS": "Q",
+}
+
+_FREQ_RE = re.compile(r"^(\d*)\s*([A-Za-z]+)(-.*)?$")
+
+
+def _translate(freq, mapping):
+    if not isinstance(freq, str):
+        return freq
+    m = _FREQ_RE.match(freq.strip())
+    if not m:
+        return freq
+    n, base, suffix = m.groups()
+    if base in mapping:
+        result = f'{n}{mapping[base]}'
+        if suffix:
+            result += suffix
+        return result
+    return freq
+
+
+# Patch to_offset (Python-level; Cython-internal callers are not
+# affected, so period_range/date_range/to_period are patched separately).
+_orig_to_offset = _freq_module.to_offset
+
+
+def _compat_to_offset(freq, *args, **kwargs):
+    return _orig_to_offset(_translate(freq, _OFFSET_MAP), *args, **kwargs)
+
+
+_freq_module.to_offset = _compat_to_offset
+
+# Patch pd.Period.__new__ (not a subclass, so type(), isinstance,
+# pickling and pydantic validation all work correctly).
+_orig_Period_new = pd.Period.__new__
+
+
+def _compat_Period_new(cls, *args, **kwargs):
+    if "freq" in kwargs:
+        kwargs["freq"] = _translate(kwargs["freq"], _PERIOD_MAP)
+    elif len(args) >= 2:
+        args = (args[0], _translate(args[1], _PERIOD_MAP)) + args[2:]
+    return _orig_Period_new(cls, *args, **kwargs)
+
+
+pd.Period.__new__ = staticmethod(_compat_Period_new)
+
+# Patch pd.period_range.
+_orig_period_range = pd.period_range
+
+
+def _compat_period_range(*args, **kwargs):
+    if "freq" in kwargs:
+        kwargs["freq"] = _translate(kwargs["freq"], _PERIOD_MAP)
+    return _orig_period_range(*args, **kwargs)
+
+
+pd.period_range = _compat_period_range
+
+# Patch pd.date_range (uses the offset naming convention).
+_orig_date_range = pd.date_range
+
+
+def _compat_date_range(*args, **kwargs):
+    if "freq" in kwargs:
+        kwargs["freq"] = _translate(kwargs["freq"], _OFFSET_MAP)
+    return _orig_date_range(*args, **kwargs)
+
+
+pd.date_range = _compat_date_range
+
+# Patch DatetimeIndex.to_period (goes through Cython to_offset
+# internally, bypassing the Python-level patch above).
+_orig_di_to_period = pd.DatetimeIndex.to_period
+
+
+def _compat_di_to_period(self, freq=None, *args, **kwargs):
+    freq = _translate(freq, _PERIOD_MAP)
+    return _orig_di_to_period(self, freq, *args, **kwargs)
+
+
+pd.DatetimeIndex.to_period = _compat_di_to_period
+
+# Restore Tick.delta property removed in pandas 3.0.
+# gluonts uses offset.delta * n to shift periods by a fixed timedelta.
+if not hasattr(_Tick, "delta"):
+    _Tick.delta = property(lambda self: pd.Timedelta(self.nanos, unit="ns"))
diff --git a/src/gluonts/transform/feature.py b/src/gluonts/transform/feature.py
index f5f519d5..13d9ea2b 100644
--- a/src/gluonts/transform/feature.py
+++ b/src/gluonts/transform/feature.py
@@ -500,7 +500,7 @@ class AddAggregateLags(MapTransformation):
             )
 
     def map_transform(self, data: DataEntry, is_train: bool) -> DataEntry:
-        assert self.base_freq == data["start"].freq
+        assert pd.tseries.frequencies.to_offset(self.base_freq) == data["start"].freq
 
         # convert to pandas Series for easier indexing and aggregation
         if is_train: