fix: benchmark timestamp ns, vectorbt trade count x2, remove fake perf numbers from README

README.md

@@ -90,15 +90,11 @@ See the [examples/](examples/) directory for complete runnable strategies:
 
 ## Performance
 
-ManifoldBT's Rust engine is orders of magnitude faster than pure-Python alternatives:
+Run the benchmark yourself:
 
-| Engine | 500K bars | 5M bars |
+```bash
-|--------|-----------|---------|
+python benchmarks/bench_vs_competitors.py --rows 500000 --runs 5
-| **ManifoldBT** | ~0.02s | ~0.15s |
+```
-| vectorbt | ~0.8s | ~8s |
-| backtrader | ~12s | ~120s+ |
-
-Run `python benchmarks/bench_vs_competitors.py` to reproduce.
 
 ## License
 

benchmarks/bench_vs_competitors.py

@@ -62,7 +62,7 @@ def bench_bt_engine(df: pd.DataFrame, n_runs: int) -> dict:
 
     # Write synthetic data to a temp parquet (manifoldbt canonical schema)
     parquet_df = pd.DataFrame({
-        "timestamp": pd.to_datetime(df["timestamp"].values, utc=True),
+        "timestamp": pd.to_datetime(df["timestamp"].values, utc=True).as_unit("ns"),
         "symbol_id": np.uint32(1),
         "open": df["open"].values,
         "high": df["high"].values,
@@ -217,7 +217,7 @@ def bench_vectorbt(df: pd.DataFrame, n_runs: int) -> dict:
         "median": np.median(times),
         "mean": np.mean(times),
         "total_return": stats.get("Total Return [%]", None),
-        "trades": stats.get("Total Trades", None),
+        "trades": int(stats.get("Total Trades", 0)) * 2,  # vbt counts round-trips; x2 for entry+exit
     }