I may have missed it but does the model require you to hold a losing trade until this trigger is met “Close the trade whenever the SPY close is higher than yesterday's high.” I.e. there is no separate stop loss mechanism?
Great strategy. I could replicate the results of the first experiment in Python with backtesting.py but when I add the market regime filter from improvement 1 the performance is lower than yours.
Hello, thanks for sharing the interesting post! I try to re-implement it, and found that the performance is worse in Improvement 2. My question is the market source is from yahoo finance or not?
Thanks for the reply! I think I've found the bug :). And another question about the "Summary of the backtest statistics", what's the difference between "Buy&Hold" and "S&P 500", I know "Buy&Hold" strategy, but how do you calculate "S&P 500" column? Really appreciate your time!
I'm not familiar with this backtesting framework you are using. So, when you call [-1], I don't know what you are referring to. I assume it's the current bar's (today) close.
Also, I don't know if you are placing orders at the day's open or the day's close. That makes quite a difference. I'm assuming orders will be placed next day's open.
Finally, there are some things funny with the logic:
- You should only buy if you are not positioned (if trade.is_long).
- I'm not using trailing stops, but a dynamic stop: if the price drops below the SMA, get out immediately.
I'm trying to reproduce your "first experiment," but I'm getting 274 trades (208 winning and 66 losing) instead of your 323 trades. The final capital is $1,547,776.79 instead of $3,092,578.
Below are the key parts of my code. Do you see any inconsistencies? Thanks.
I managed to identify the issue—it was the data source. I was using the IB database and found multiple errors. After switching to Yahoo Finance, everything worked correctly.
The only significantly different parameter compared to Experiment 1 (with QQQ) is the Sharpe Ratio. I got 0.67, whereas yours was 1.83. I used a risk-free rate of 0.
Great to hear you were able to replicate the results!
Regarding the Sharpe ratio, the difference is that you are considering days where you had no position (hence no risk) in your computation. That doesn't make sense. We should exclude these days. One thing is to take a risk on a given day and have zero return (this day should be included). Another completely different thing is to not take risks (no position) on a given day (this day should be excluded).
I may have missed it but does the model require you to hold a losing trade until this trigger is met “Close the trade whenever the SPY close is higher than yesterday's high.” I.e. there is no separate stop loss mechanism?
In the "Improvement 2" section, we hold the trades until either of the following is triggered:
- Close the trade whenever the price is higher than yesterday's high (same as before);
- Close the trade whenever the price is lower than the 300-SMA (new condition).
This is more than enough. The first event (price closing higher than yesterday's high) is pretty frequent: it occurs on average once every 3 days :)
Hi there, have you assumed any commissions?
Not in this exercise... I assumed slippage only (1 basis point against all transactions)
Hi, when you refer to the rolling high over 10 days in bullet 3, do you mean the mean of high over a 10 day window?
Hi Mike! No, it’s rolling high. In python:
df[(asset, 'Band High')] = t[(asset, 'High')].rolling(window=10).max()
well done!
Great strategy. I could replicate the results of the first experiment in Python with backtesting.py but when I add the market regime filter from improvement 1 the performance is lower than yours.
It's just additional
- go long when close > sma300
- exit trade when close < sma300
right?
Hi, thanks! Not quite... the improvement 1 means we only take trades if close > sma300.
If that's the case, then follow the previously seen rules. Otherwise, don't bother and skip the day. Cheers!
Thank you for the fast response. My result are still slightly lower but that might be caused by different backtesting algo and data source.
Could you please state the rules for improvement 3 (Long & Short) as well?
I don't understand why there's such a long idle period in the beginning. Shouldn't the algo trade the Short ETF during the dotcom crash?
Do you enter the trade at the close, or the next day's open?
Hi! Always next day's open
Consider to apply this model on 20 other equity indices to see if this is a robust one.
Good idea! Will do that, thx!
Hello, thanks for sharing the interesting post! I try to re-implement it, and found that the performance is worse in Improvement 2. My question is the market source is from yahoo finance or not?
I'm currently using Norgate data and Sharadar Core US Equities Bundle.
But I've used free datasets such as Stooq and Yahoo Finance in the past.
The problem with free datasets is that the quality is extremely low.
They frequently forget to adjust for corporate actions, there are missing values, discontinuities, etc...
Thanks for the reply! I think I've found the bug :). And another question about the "Summary of the backtest statistics", what's the difference between "Buy&Hold" and "S&P 500", I know "Buy&Hold" strategy, but how do you calculate "S&P 500" column? Really appreciate your time!
Thanks! If the target asset is QQQ, for instance, Buy&Hold means B&H QQQ.
Cheers
Thanks for the explanation!
I've tried to reproduce your results. But I got much less trades in the similar timeframe.
Here are my snippets:
def high_low_mean(high: pd.Series, low: pd.Series, interval: int) -> pd.Series:
return (high - low).rolling(interval).mean()
def lower_band(high: pd.Series, hl: pd.Series, intervall:int, factor: float) -> pd.Series:
return high.rolling(intervall).high() - factor * hl
stock["ibs"] = (stock.Close - stock.Low) / (stock.High - stock.Low)
The Roule is to buy on next days open on following condition (-1 means current day):
if (
(_close < self.lower_band[-1])
and (self.ibs[-1] < (self.ibs_low))
and (_low > self.sma[-1])
):
self.buy(sl=self.sma[-1])
Trade management:
if trade.is_long:
# update trailing stop
trade.sl = max(trade.sl, self.sma[-1])
# close on next day open
if _close > self.data.High[-2]:
trade.close()
Hi @Markus,
I'm not familiar with this backtesting framework you are using. So, when you call [-1], I don't know what you are referring to. I assume it's the current bar's (today) close.
Also, I don't know if you are placing orders at the day's open or the day's close. That makes quite a difference. I'm assuming orders will be placed next day's open.
Finally, there are some things funny with the logic:
- You should only buy if you are not positioned (if trade.is_long).
- I'm not using trailing stops, but a dynamic stop: if the price drops below the SMA, get out immediately.
Hope it helps!
I'm trying to reproduce your "first experiment," but I'm getting 274 trades (208 winning and 66 losing) instead of your 323 trades. The final capital is $1,547,776.79 instead of $3,092,578.
Below are the key parts of my code. Do you see any inconsistencies? Thanks.
df['range'] = df['High'] - df['Low']
df['mean_range_25'] = df['range'].rolling(25).mean()
df['IBS'] = (df['Close'] - df['Low']) / df['range']
df['rolling_high_10'] = df['High'].rolling(10).max()
df['lower_band'] = df['rolling_high_10'] - (2.5 * df['mean_range_25'])
# BUY logic
if not in_position:
if (close_today < lower_band_today) and (ibs_today < 0.3):
# Buy as many shares as possible
shares_held = int(capital // close_today)
cost = shares_held * close_today
capital -= cost
in_position = True
entry_date = today
entry_price = close_today
entry_shares = shares_held
total_equity = shares_held * close_today
# SELL logic
else:
if close_today > high_yesterday:
exit_date = today
exit_price = close_today
trade_pnl = entry_shares * (exit_price - entry_price)
trade_duration = (exit_date - entry_date).days
I managed to identify the issue—it was the data source. I was using the IB database and found multiple errors. After switching to Yahoo Finance, everything worked correctly.
______________________________________________________________________
Performance Summary
Period: March 10, 1999 – May 17, 2024
Initial Capital: $100,000.00
Final Capital: $3,073,003.54
Total Return: 2973.00%
Sharpe Ratio: 0.67
Max Drawdown: -24.57%
Average Drawdown: -3.54%
Trade Statistics
# Trades Trades / Year Avg. Return / Trade [%] Best / Worst Trade [%] Win/Loss Ratio Payoff Ratio CPC Index Expectancy [$]
All Trades 323 12.87 1.12 16.84 / -18.14 2.67 1.05 2.64
Winning Trades 235 9.36 2.38 16.84 / 0.03 NaN NaN NaN
Losing Trades 88 3.51 -2.27 0.00 / -18.14 0.00 NaN NaN
______________________________________________________________________
The only significantly different parameter compared to Experiment 1 (with QQQ) is the Sharpe Ratio. I got 0.67, whereas yours was 1.83. I used a risk-free rate of 0.
Here is the code:
______________________________________________________________________
def calculate_sharpe_ratio(equity_series, periods_per_year=252, risk_free_rate=0):
returns = equity_series.pct_change().dropna()
excess_daily_returns = returns - (risk_free_rate / periods_per_year)
std_daily = excess_daily_returns.std()
if std_daily == 0:
return 0.0
sharpe = (excess_daily_returns.mean() / std_daily) * np.sqrt(periods_per_year)
return sharpe
Great to hear you were able to replicate the results!
Regarding the Sharpe ratio, the difference is that you are considering days where you had no position (hence no risk) in your computation. That doesn't make sense. We should exclude these days. One thing is to take a risk on a given day and have zero return (this day should be included). Another completely different thing is to not take risks (no position) on a given day (this day should be excluded).