Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(shadow-account): seed Wilder RSI with a simple average, not the first observation

Lanre Shittu85c5079f9dc0
diff --git a/agent/src/shadow_account/extractor.py b/agent/src/shadow_account/extractor.py
index dbbb8b7e6..dd32b5ba9 100644
--- a/agent/src/shadow_account/extractor.py
+++ b/agent/src/shadow_account/extractor.py
@@ -147,8 +147,50 @@ def extract_shadow_profile(
# ---------------- Feature engineering ----------------
+def _wilder_average(values: pd.Series, period: int) -> pd.Series:
+ """True Wilder recursive smoothing: SMA-seeded, then exponential recursion.
+
+ Wilder's original formula seeds the first average as a simple mean of the
+ first ``period`` valid values, then recurses as
+ ``avg[t] = (avg[t-1] * (period - 1) + values[t]) / period`` for every bar
+ after that. This is *not* the same as
+ ``values.ewm(alpha=1/period, adjust=False).mean()``, which seeds its
+ recursion from the single first observation instead of that average —
+ measurably further from Wilder during the warmup-adjacent bars, not
+ closer, even though the two converge for data far enough past warmup.
+
+ Vectorized as an ``ewm`` over the seed plus the remaining values, which is
+ mathematically identical to the explicit recursion: seeding an
+ ``adjust=False`` ewm with the average as its first "observation" and
+ dropping that duplicate first row reproduces the same recursive update at
+ every following step.
+
+ Args:
+ values: Series to smooth (e.g. per-bar gains or losses); may lead
+ with NaN (as ``close.diff()`` does).
+ period: Wilder lookback period.
+
+ Returns:
+ A series aligned with ``values``, NaN before the seed index.
+ """
+ result = pd.Series(np.nan, index=values.index, dtype=float)
+ valid = values.dropna()
+ if len(valid) < period:
+ return result
+
+ seed = valid.iloc[:period].mean()
+ result.loc[valid.index[period - 1]] = seed
+
+ remaining = valid.iloc[period:]
+ if len(remaining) > 0:
+ tail = pd.concat([pd.Series([seed]), remaining])
+ smoothed_tail = tail.ewm(alpha=1 / period, adjust=False).mean().iloc[1:]
+ result.loc[remaining.index] = smoothed_tail.to_numpy()
+ return result
+
+
def _compute_rsi(close: pd.Series, period: int = _RSI_PERIOD) -> pd.Series:
- """Causal Wilder-EWM RSI.
+ """Causal Wilder RSI.
Mirrors the shape of ``compute_rsi`` in
``agent/src/skills/technical-basic/example_signal_engine.py:13`` — that
@@ -166,8 +208,8 @@ def _compute_rsi(close: pd.Series, period: int = _RSI_PERIOD) -> pd.Series:
delta = close.diff()
gain = delta.clip(lower=0)
loss = (-delta).clip(lower=0)
- avg_gain = gain.ewm(alpha=1 / period, adjust=False, min_periods=period).mean()
- avg_loss = loss.ewm(alpha=1 / period, adjust=False, min_periods=period).mean()
+ avg_gain = _wilder_average(gain, period)
+ avg_loss = _wilder_average(loss, period)
rs = avg_gain / avg_loss
return 100 - 100 / (1 + rs)
diff --git a/agent/src/shadow_account/scanner.py b/agent/src/shadow_account/scanner.py
index d3bb3ba6e..88e7aa546 100644
--- a/agent/src/shadow_account/scanner.py
+++ b/agent/src/shadow_account/scanner.py
@@ -9,6 +9,7 @@ from __future__ import annotations
from datetime import date, datetime
from typing import Any, Callable
+import numpy as np
import pandas as pd
from src.shadow_account.backtester import _LIQUID_BASKETS, SUPPORTED_MARKETS
@@ -200,8 +201,33 @@ def _compute_features(bars: pd.DataFrame, rule: ShadowRule) -> dict[str, float |
return features
+def _wilder_average(values: pd.Series, period: int) -> pd.Series:
+ """True Wilder recursive smoothing: SMA-seeded, then exponential recursion.
+
+ Kept byte-identical to ``_wilder_average`` in ``extractor.py`` — see that
+ copy's docstring for the full rationale. Duplicated rather than imported
+ for the same reason ``_compute_rsi`` already was: this module's
+ ``_compute_rsi`` must evaluate the same RSI that produced the extracted
+ ``entry_rsi14`` bounds, so the two copies cannot drift apart.
+ """
+ result = pd.Series(np.nan, index=values.index, dtype=float)
+ valid = values.dropna()
+ if len(valid) < period:
+ return result
+
+ seed = valid.iloc[:period].mean()
+ result.loc[valid.index[period - 1]] = seed
+
+ remaining = valid.iloc[period:]
+ if len(remaining) > 0:
+ tail = pd.concat([pd.Series([seed]), remaining])
+ smoothed_tail = tail.ewm(alpha=1 / period, adjust=False).mean().iloc[1:]
+ result.loc[remaining.index] = smoothed_tail.to_numpy()
+ return result
+
+
def _compute_rsi(close: pd.Series, period: int = _RSI_PERIOD) -> pd.Series:
- """Causal Wilder-EWM RSI.
+ """Causal Wilder RSI.
Mirrors ``_compute_rsi`` in ``extractor.py`` so the scanner evaluates the
same RSI that produced the extracted ``entry_rsi14`` bounds. Causal by
@@ -210,8 +236,8 @@ def _compute_rsi(close: pd.Series, period: int = _RSI_PERIOD) -> pd.Series:
delta = close.diff()
gain = delta.clip(lower=0)
loss = (-delta).clip(lower=0)
- avg_gain = gain.ewm(alpha=1 / period, min_periods=period).mean()
- avg_loss = loss.ewm(alpha=1 / period, min_periods=period).mean()
+ avg_gain = _wilder_average(gain, period)
+ avg_loss = _wilder_average(loss, period)
rs = avg_gain / avg_loss
return 100 - 100 / (1 + rs)
diff --git a/agent/tests/test_shadow_account.py b/agent/tests/test_shadow_account.py
index a6820e9b2..7f0170c82 100644
--- a/agent/tests/test_shadow_account.py
+++ b/agent/tests/test_shadow_account.py
@@ -1214,11 +1214,15 @@ def test_compute_rsi_is_causal_and_bounded() -> None:
@pytest.mark.unit
def test_compute_rsi_matches_wilder_recursive_smoothing() -> None:
"""On a mixed gain/loss series, must match true Wilder recursive
- smoothing, not pandas' default weight-normalized ewm (which diverges
- materially from Wilder on short warmup-adjacent windows)."""
+ smoothing: the first average is a simple mean of the first `period`
+ gains/losses (the seed), and every bar after that applies the recursive
+ update ``avg[t] = (avg[t-1] * (period - 1) + x[t]) / period``. Seeding
+ from the single first observation instead (pandas' ewm with
+ adjust=False, taken alone) diverges materially from this on the bars
+ right after warmup, even though the two converge much later."""
closes = [
100.0, 101.0, 99.5, 102.0, 101.0, 103.0, 100.0, 104.0, 101.5, 105.0,
- 102.0, 106.0, 103.0, 107.0, 104.5,
+ 102.0, 106.0, 103.0, 107.0, 104.5, 106.0, 108.0, 103.5, 107.0, 109.0,
]
close = pd.Series(closes)
@@ -1226,16 +1230,39 @@ def test_compute_rsi_matches_wilder_recursive_smoothing() -> None:
gains = [max(d, 0.0) for d in deltas]
losses = [max(-d, 0.0) for d in deltas]
period = 14
- alpha = 1 / period
- avg_gain, avg_loss = gains[0], losses[0]
- for g, loss_val in zip(gains[1:], losses[1:]):
- avg_gain = alpha * g + (1 - alpha) * avg_gain
- avg_loss = alpha * loss_val + (1 - alpha) * avg_loss
+
+ avg_gain = sum(gains[:period]) / period
+ avg_loss = sum(losses[:period]) / period
+ for g, loss_val in zip(gains[period:], losses[period:]):
+ avg_gain = (avg_gain * (period - 1) + g) / period
+ avg_loss = (avg_loss * (period - 1) + loss_val) / period
expected = 100 - 100 / (1 + avg_gain / avg_loss)
assert float(_compute_rsi(close, period=period).iloc[-1]) == pytest.approx(expected)
+def test_compute_rsi_seed_is_a_simple_average_not_first_value() -> None:
+ """At exactly `period` valid deltas (the boundary case), the seed value
+ itself is the whole answer: a simple mean of every gain/loss so far, not
+ the single-observation seed pandas' ewm(adjust=False) would use."""
+ closes = [
+ 100.0, 101.0, 99.5, 102.0, 101.0, 103.0, 100.0, 104.0, 101.5, 105.0,
+ 102.0, 106.0, 103.0, 107.0, 104.5,
+ ]
+ close = pd.Series(closes)
+
+ deltas = [closes[i] - closes[i - 1] for i in range(1, len(closes))]
+ gains = [max(d, 0.0) for d in deltas]
+ losses = [max(-d, 0.0) for d in deltas]
+ period = 14
+ assert len(gains) == period
+
+ expected = 100 - 100 / (1 + (sum(gains) / period) / (sum(losses) / period))
+
+ assert float(_compute_rsi(close, period=period).iloc[-1]) == pytest.approx(expected)
+
+
@pytest.mark.unit
def test_price_features_as_of_reads_only_past_bars() -> None:
dates = [f"2026-02-{d:02d}" for d in range(1, 21)]