Reproduire l’analyse exploratoire de research.ipynb avec les données natives QuantConnect.
# Setup QuantBook
from AlgorithmImports import *
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')
plt.style.use('seaborn-v0_8-darkgrid')
plt.rcParams['figure.figsize'] = (14, 6)
qb = QuantBook()
print("QuantBook initialise.")QuantBook initialise.4 paires actuelles + 2 candidates supplementaires.
fx_pairs = ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD', 'GBPUSD', 'NZDUSD']
symbols = {}
for pair in fx_pairs:
try:
symbols[pair] = qb.add_forex(pair, Resolution.DAILY).symbol
except:
print(f"Warning: {pair} non disponible")
start = datetime(2010, 1, 1)
end = datetime(2026, 1, 1)
history = qb.history(list(symbols.values()), start, end, Resolution.DAILY)
closes = history['close'].unstack(level=0)
stm = {str(v): k for k, v in symbols.items()}
closes.columns = [stm.get(str(c), str(c)) for c in closes.columns]
closes = closes.dropna()
# QC History retourne les bars daily a 16:00 (cloture marche), pas a minuit.
# resample('MS') etiquette les bacs a minuit -> le test d'appartenance mensuel
# (date in closes.index) ne matcherait jamais l'index daily -> on normalise a minuit.
closes.index = closes.index.normalize()
print(f"Periode: {closes.index[0].date()} a {closes.index[-1].date()}")
print(f"Donnees: {len(closes)} jours de trading")
print(f"Paires: {list(closes.columns)}")
returns_df = closes.pct_change()Periode: 2010-01-03 a 2025-12-31
Donnees: 4996 jours de trading
Paires: ['AUDUSD', 'EURUSD', 'GBPUSD', 'NZDUSD', 'USDCAD', 'USDJPY']Note sur les paires USD/XXX :
Pour USDJPY et USDCAD, un prix qui monte = USD fort = devise locale faible. Pour mesurer le momentum de la devise (pas du USD), on inverse le return.
# Statistiques par paire
print(f"{'Paire':<10} {'Rend. Ann.':>12} {'Volatilite':>12} {'Sharpe':>8}")
print("-" * 45)
for pair in closes.columns:
ret = (closes[pair].iloc[-1] / closes[pair].iloc[0]) ** (252 / len(closes)) - 1
vol = returns_df[pair].std() * np.sqrt(252)
sharpe = ret / vol if vol > 0 else 0
print(f"{pair:<10} {ret:>11.1%} {vol:>11.1%} {sharpe:>7.2f}")Paire Rend. Ann. Volatilite Sharpe
---------------------------------------------
AUDUSD -1.5% 9.6% -0.15
EURUSD -1.0% 7.8% -0.13
GBPUSD -0.9% 8.0% -0.11
NZDUSD -1.1% 9.9% -0.11
USDCAD 1.3% 6.8% 0.20
USDJPY 2.6% 8.4% 0.31def fx_momentum_score(closes, pair, lookback_short=21, lookback_long=126,
w_short=0.7, w_long=0.3, risk_adjusted=True,
skip_month=True):
"""Compute composite FX momentum score."""
prices = closes[pair]
# Invert USD/XXX pairs (USD strength = negative momentum for the currency)
invert = pair.startswith('USD')
if skip_month:
# Skip last 21 days (Jegadeesh)
ret_long = prices.shift(21) / prices.shift(lookback_long) - 1
else:
ret_long = prices / prices.shift(lookback_long) - 1
ret_short = prices / prices.shift(lookback_short) - 1
if invert:
ret_long = -ret_long
ret_short = -ret_short
# Composite score
score = w_short * ret_short + w_long * ret_long
if risk_adjusted:
vol = returns_df[pair].rolling(63).std() * np.sqrt(252)
score = score / vol.clip(lower=0.01)
return score
def backtest_fx_momentum(closes, universe, top_n=2, lookback_short=21,
lookback_long=126, risk_adjusted=True,
skip_month=True):
"""Backtest FX momentum strategy."""
returns_df = closes.pct_change()
# Compute scores for all pairs
scores = pd.DataFrame(index=closes.index)
for pair in universe:
if pair in closes.columns:
scores[pair] = fx_momentum_score(
closes, pair, lookback_short, lookback_long,
risk_adjusted=risk_adjusted, skip_month=skip_month
)
# Monthly rebalance
monthly = closes.resample('MS').first().index
start_idx = max(lookback_long + 21, 63) + 1
port_ret = pd.Series(0.0, index=closes.index)
current_pos = []
for date in monthly:
if date not in closes.index:
continue
if date < closes.index[start_idx]:
continue
s = scores.loc[date].dropna()
if len(s) == 0:
continue
# Top-N with positive IR
top = s.nlargest(top_n)
current_pos = [p for p in top.index if top[p] > 0]
# Actually compute returns with proper rebalancing
positions = {}
for date in monthly:
if date not in closes.index or date < closes.index[start_idx]:
continue
s = scores.loc[date].dropna()
if len(s) == 0:
positions[date] = []
continue
top = s.nlargest(top_n)
positions[date] = [p for p in top.index if top[p] > 0]
current_pos = []
weight = 0.5 # 50% per position
for i in range(start_idx, len(closes)):
date = closes.index[i]
# Check if rebalance day
if date in positions:
current_pos = positions[date]
if len(current_pos) > 0:
w = 1.0 / max(len(current_pos), 1)
for pair in current_pos:
if pair in returns_df.columns:
# For USD/XXX pairs, we need to invert
r = returns_df[pair].iloc[i]
if pair.startswith('USD'):
r = -r
port_ret.iloc[i] += w * r
port_ret = port_ret[port_ret.index >= closes.index[start_idx]]
vals = (1 + port_ret).cumprod()
total = vals.iloc[-1] - 1
years = len(port_ret) / 252
cagr = (1 + total) ** (1 / years) - 1 if years > 0 else 0
vol = port_ret.std() * np.sqrt(252)
sharpe = (cagr - 0.03) / vol if vol > 0.001 else 0
max_dd = ((vals - vals.expanding().max()) / vals.expanding().max()).min()
return {'sharpe': sharpe, 'cagr': cagr, 'max_dd': max_dd, 'vol': vol, 'cum': vals}
print("Fonctions definies.")Fonctions definies.Règle #1 du backlog attendait « risk-adjusted momentum > raw momentum ». La mesure FX dit l’inverse (cf. Verdict H1 ci-dessous).
base_universe = ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD']
print(f"{'Momentum Type':<25} {'Sharpe':>8} {'CAGR':>8} {'MaxDD':>8}")
print("-" * 50)
results_mom = {}
for name, ra in [('Pure momentum', False), ('Risk-adjusted (actuel)', True)]:
r = backtest_fx_momentum(closes, base_universe, risk_adjusted=ra)
results_mom[name] = r
print(f"{name:<25} {r['sharpe']:>8.3f} {r['cagr']:>7.1%} {r['max_dd']:>7.1%}")Momentum Type Sharpe CAGR MaxDD
--------------------------------------------------
Pure momentum -0.719 -1.2% -33.2%
Risk-adjusted (actuel) -0.744 -1.3% -35.7%research.ipynb etait inconclusif sur FX. La re-exec tranche : le risk-adjustment nuit sur FX G10 — Pure momentum Sharpe -0.719 (CAGR -1.2%, MaxDD -33.2%) surperforme le risk-adjusted -0.744 (-1.3%, -35.7%). Ecart avec la Règle #1 documente au tableau §9.
print(f"{'Skip-Month':<25} {'Sharpe':>8} {'CAGR':>8} {'MaxDD':>8}")
print("-" * 50)
results_skip = {}
for name, skip in [('Sans skip (J-126 a J-0)', False), ('Avec skip (J-126 a J-21)', True)]:
r = backtest_fx_momentum(closes, base_universe, skip_month=skip)
results_skip[name] = r
print(f"{name:<25} {r['sharpe']:>8.3f} {r['cagr']:>7.1%} {r['max_dd']:>7.1%}")Skip-Month Sharpe CAGR MaxDD
--------------------------------------------------
Sans skip (J-126 a J-0) -0.767 -1.5% -37.6%
Avec skip (J-126 a J-21) -0.744 -1.3% -35.7%Règle #2 du backlog: skip-month confirme pour equities (Jegadeesh 1990). Sur FX, la mesure donne un effet faible et en demi-teinte : Avec skip Sharpe -0.744 (CAGR -1.3%, MaxDD -35.7%) vs Sans skip -0.767 (-1.5%, -37.6%). Le skip attenu legerement la perte (Δ +0.023 Sharpe) mais les deux configurations restent negatives — l’effet skip ne sauve pas la strategie, il la rend un peu moins mauvaise.
print(f"{'Lookback':<20} {'Sharpe':>8} {'CAGR':>8} {'MaxDD':>8}")
print("-" * 45)
results_lb = {}
for lb_name, lb_d in [('21d (1m)', 21), ('63d (3m)', 63), ('126d (6m, actuel)', 126), ('252d (12m)', 252)]:
r = backtest_fx_momentum(closes, base_universe, lookback_long=lb_d)
results_lb[lb_name] = r
print(f"{lb_name:<20} {r['sharpe']:>8.3f} {r['cagr']:>7.1%} {r['max_dd']:>7.1%}")Lookback Sharpe CAGR MaxDD
---------------------------------------------
21d (1m) -0.634 -0.8% -27.9%
63d (3m) -0.778 -1.5% -37.2%
126d (6m, actuel) -0.744 -1.3% -35.7%
252d (12m) -0.539 -0.0% -17.7%126d (6 mois) est le lookback long actuel. FX momentum est typiquement plus court que equity momentum (1-3 mois vs 12 mois).
print(f"{'Top-N':<15} {'Sharpe':>8} {'CAGR':>8} {'MaxDD':>8}")
print("-" * 40)
results_topn = {}
for n in [1, 2, 3]:
r = backtest_fx_momentum(closes, base_universe, top_n=n)
name = f'Top-{n}'
results_topn[name] = r
print(f"{name:<15} {r['sharpe']:>8.3f} {r['cagr']:>7.1%} {r['max_dd']:>7.1%}")Top-N Sharpe CAGR MaxDD
----------------------------------------
Top-1 -0.713 -1.4% -38.5%
Top-2 -0.744 -1.3% -35.7%
Top-3 -0.738 -1.2% -33.0%Top-2 (actuel) = 50% par position. Top-1 = 100% concentration. Top-3 = 33% mais sur 4 paires = presque equal weight.
universes = {
'4 paires (actuel)': ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD'],
'+ GBPUSD': ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD', 'GBPUSD'],
'+ NZDUSD': ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD', 'NZDUSD'],
'6 paires': ['EURUSD', 'AUDUSD', 'USDJPY', 'USDCAD', 'GBPUSD', 'NZDUSD'],
}
print(f"{'Univers':<25} {'Sharpe':>8} {'CAGR':>8} {'MaxDD':>8}")
print("-" * 50)
results_univ = {}
for name, univ in universes.items():
avail = [p for p in univ if p in closes.columns]
r = backtest_fx_momentum(closes, avail)
results_univ[name] = r
print(f"{name:<25} {r['sharpe']:>8.3f} {r['cagr']:>7.1%} {r['max_dd']:>7.1%}")Univers Sharpe CAGR MaxDD
--------------------------------------------------
4 paires (actuel) -0.744 -1.3% -35.7%
+ GBPUSD -0.722 -1.4% -37.2%
+ NZDUSD -0.785 -1.8% -42.0%
6 paires -0.727 -1.6% -40.3%iter5 a teste 6 paires + top-3 + USD trend filter -> Sharpe -0.849. Plus de paires ne sauve pas le momentum FX si le problème est structurel.
fig, axes = plt.subplots(2, 2, figsize=(16, 10))
ax = axes[0, 0]
for name, r in results_mom.items():
ax.plot(r['cum'].values, label=f"{name} (S={r['sharpe']:.2f})", linewidth=1.5)
ax.set_title('H1: Pure vs Risk-adjusted momentum', fontweight='bold')
ax.legend(fontsize=7)
ax.grid(True, alpha=0.3)
ax = axes[0, 1]
for name, r in results_lb.items():
ax.plot(r['cum'].values, label=f"{name} (S={r['sharpe']:.2f})", linewidth=1.5)
ax.set_title('H3: Lookback period', fontweight='bold')
ax.legend(fontsize=7)
ax.grid(True, alpha=0.3)
ax = axes[1, 0]
for name, r in results_topn.items():
ax.plot(r['cum'].values, label=f"{name} (S={r['sharpe']:.2f})", linewidth=1.5)
ax.set_title('H4: Top-N positions', fontweight='bold')
ax.legend(fontsize=8)
ax.grid(True, alpha=0.3)
ax = axes[1, 1]
for name, r in results_univ.items():
ax.plot(r['cum'].values, label=f"{name} (S={r['sharpe']:.2f})", linewidth=1.5)
ax.set_title('H5: Universe expansion', fontweight='bold')
ax.legend(fontsize=7)
ax.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig('forexcarry_quantbook_analysis.png', dpi=150, bbox_inches='tight')
plt.show()
| Hypothese | Resultat QuantBook (tous negatifs) |
|---|---|
| H1 Risk-adjusted | Pure -0.719 surperforme Risk-adjusted -0.744 (contredit Règle #1) |
| H2 Skip-month | Avec skip -0.744 vs Sans skip -0.767 (effet faible, les 2 negatifs) |
| H3 Lookback | 252d (12m) le moins pire : -0.539 ; 21d -0.634 ; 126d -0.744 ; 63d -0.778 |
| H4 Top-N | Top-1 le moins pire : -0.713 ; Top-3 -0.738 ; Top-2 -0.744 |
| H5 Universe | +GBPUSD -0.722 ; 6 paires -0.727 ; 4 paires -0.744 ; +NZDUSD -0.785 |
Verdict : NO BEATS — aucune configuration ne produit un Sharpe positif. Données FX issues de yfinance (backend LEAN, cf. corps de PR) ; pas de cross-check yfinance independant possible (yfinance est la source).
FX momentum G10 genere ~0.7-1.5% CAGR absolu, mais le taux sans risque est de 2.5%-5.5% sur 2015-2026. Le Sharpe negatif est donc structurel: le momentum FX ne compense pas le cout d’opportunite du cash.
Menkhoff et al. (2012) documentent l’affaiblissement du FX momentum post-2008, lie a l’intervention massive des banques centrales.