Original paper
Abstract
Recent research documents that commodities are good diversifiers in traditional investment portfolios: overall portfolio risk is reduced while less than proportional return is sacrificed. These studies generally find a relatively high volatility in commodity returns, which implies a huge potential for tactical strategies. In this paper we investigate timing strategies with commodity futures using factors directly related to the stance of the business cycle, the monetary environment and the sentiment of the market. We use a dynamic model selection procedure in the spirit of the recursive modeling approach of Pesaran and Timmermann [1995]. However, instead of using in-sample model selection criteria, we build on the extensions of Bauer, Derwall and Molenaar [2004] by introducing an out-of-sample model training period to select optimal models. The best models from this training period are used to generate forecasts in a subsequent trading period. Our results show that the variation in commodity future returns is sufficiently predictable to be exploited by a realistic timing strategy.
Keywords:Â Commodity futures, market timing
Trading rules
- Use GSCI index (or GSG ETF) for trading.
- Apply a three-phase approach: estimation in-sample, training, and trading.
- Utilize 15 forecasting variables in the in-sample period.
- Restrict explanatory variables to between 0 and 5 (4,944 possible models).
- Estimate parameters using OLS regression.
- Generate monthly signals during a 24-month training period.
- Buy GSCI-index futures for positive signals, sell for negative signals.
- Evaluate models considering their realized information ratios and transaction costs.
- Select model with highest realized information ratio for next month’s forecast.
- Implement a structure of 60 months for initial estimation, 24 months for training, and a 1-month trading interval.
- Conduct monthly re-calibration and model selections.
- Rebalance every month.
Python code
Backtrader
import backtrader as bt
import pandas as pd
import numpy as np
import statsmodels.api as sm
from itertools import combinations
class GSCIIndexStrategy(bt.Strategy):
params = (
("insample_window", 60),
("training_period", 24),
("trading_period", 1),
("num_variables", 15),
("min_variables", 0),
("max_variables", 5),
)
def __init__(self):
self.month_counter = 0
self.best_model = None
def next(self):
self.month_counter += 1
if self.month_counter % self.params.trading_period == 0:
self.rebalance()
def rebalance(self):
if self.data.datetime.date(0).month != self.data.datetime.date(-1).month:
insample_data = self.data.get(size=self.params.insample_window)
self.best_model = self.select_best_model(insample_data)
training_data = self.data.get(size=self.params.training_period)
signal = self.generate_signal(self.best_model, training_data)
if signal > 0:
self.buy()
elif signal < 0:
self.sell()
def select_best_model(self, data):
best_ir = -np.inf
best_model = None
for num_vars in range(self.params.min_variables, self.params.max_variables + 1):
for combo in combinations(range(self.params.num_variables), num_vars):
X = data[:, combo]
X = sm.add_constant(X)
y = data[:, -1]
model = sm.OLS(y, X).fit()
ir = model.tvalues[-1] / model.bse[-1]
if ir > best_ir:
best_ir = ir
best_model = model
return best_model
def generate_signal(self, model, data):
X = data[:, model.params.index[1:]]
X = sm.add_constant(X)
signal = model.predict(X[-1])
return signal
if __name__ == "__main__":
cerebro = bt.Cerebro()
cerebro.addstrategy(GSCIIndexStrategy)
data = bt.feeds.YahooFinanceData(dataname="GSG", fromdate="1990-01-01", todate="2021-09-30")
cerebro.adddata(data)
cerebro.broker.setcash(100000.0)
cerebro.broker.setcommission(commission=0.002)
print("Starting Portfolio Value:", cerebro.broker.getvalue())
cerebro.run()
print("Final Portfolio Value:", cerebro.broker.getvalue())Please note that this code is a simplified version of the strategy, and you may need to adjust or add features depending on your specific data and requirements. You will also need to provide the appropriate data containing the forecasting variables for the strategy to work correctly.