Keep is_buyer_maker aligned with its trade when TradesData sorts and de-duplicates

finmlkit/bar/data_model.py

@@ -162,6 +162,11 @@ def __init__(self,
         self._data = pd.DataFrame({'timestamp': ts, 'price': px, 'amount': qty, 'id': id})
 
         self.is_buyer_maker = is_buyer_maker
+        if is_buyer_maker is not None and preprocess:
+            if len(is_buyer_maker) != len(self._data):
+                raise ValueError("is_buyer_maker must have the same length as the trades")
+            # Carried as a column so that sorting and de-duplicating re-order it together with its trade
+            self._data['is_buyer_maker'] = is_buyer_maker
         if side is not None:
             self._data['side'] = side
         self._orig_timestamp_unit = timestamp_unit if timestamp_unit else self._infer_timestamp_unit()
@@ -329,6 +334,10 @@ def _merge_trades(self):
         """
         logger.info('Merging split trades (same timestamps) on same price level...')
 
+        if 'is_buyer_maker' in self._data.columns:
+            # re-ordered together with the other columns by _sort_trades
+            self.is_buyer_maker = self._data['is_buyer_maker'].to_numpy(dtype=np.bool_)
+
         ts, px, am, side = merge_split_trades(
             self.data['timestamp'].values.astype(np.int64),
             self.data['price'].values.astype(np.float64),

tests/bars/test_trades_data_side_alignment.py

@@ -0,0 +1,29 @@
+import numpy as np
+
+from finmlkit.bar.data_model import TradesData
+
+
+def test_side_stays_with_its_trade_when_input_is_not_in_id_order():
+    """Binance dumps are not always in id order; sorting must carry is_buyer_maker along with price/time/amount."""
+    n = 1000
+    ids = np.arange(1, n + 1, dtype=np.int64)
+    ts = (1_700_000_000_000 + ids * 5).astype(np.int64)            # ms, one trade per timestamp: nothing merges
+    px = 100.0 + ids * 0.01
+    qty = np.ones(n)
+    buyer_maker = (ids % 3 == 0)                                    # a side that is a function of the trade
+    order = np.random.default_rng(0).permutation(n)                 # shuffled, as in an out-of-order dump
+    td = TradesData(ts[order], px[order], qty[order], id=ids[order], is_buyer_maker=buyer_maker[order],
+                    timestamp_unit="ms", preprocess=True)
+    out = td.data
+    trade_id = np.rint((out["price"].to_numpy() - 100.0) / 0.01).astype(np.int64)
+    expected = np.where(trade_id % 3 == 0, -1, 1)                   # buyer maker => seller aggressor => -1
+    assert (out["side"].to_numpy() == expected).all()
+
+
+def test_sorted_input_is_unchanged():
+    n = 50
+    ids = np.arange(1, n + 1, dtype=np.int64)
+    ts = (1_700_000_000_000 + ids * 5).astype(np.int64)
+    td = TradesData(ts, 100.0 + ids, np.ones(n), id=ids, is_buyer_maker=(ids % 2 == 0),
+                    timestamp_unit="ms", preprocess=True)
+    assert (td.data["side"].to_numpy() == np.where(ids % 2 == 0, -1, 1)).all()