Why choose IGM as the benchmark? Choose AAPL as the benchmark and see that one would have done better to just buy and hold. Buy and hold has about the same volatility (.34) higher Sharpe ratio (1.02) but does have greater drawdown (60.82%). Here's the same algo but with the benchmark set to AAPL.

Thank you for pointing out Dan. I too noticed the drastic algorithm under-performance when setting Apple as the benchmark. Any insight into why trend surfing fell short of just holding long?

@Conrad, @Dan is right in his observation, and makes the right argument.

The reason why your trade slicing produces less than the simple buy & hold is that you get whipsawed at lot due to your lagging response to price change. The positive trades are not sufficient or large enough to compensate for this deficiency. A way of saying you need a better strategy design.

Here is the equation to follow: Σ(q∙Δp) > B&H = q_0∙[p(T) – p_0], which says that the sum of the sequence of trades (P&L) need to beat the buy & hold scenario which is simply making the initial investment (F_0 / p_0 = q_0) over the time interval (T - t_0).

@Conrad, what you presented is an all too frequent example of the sum of the parts did not exceed the whole. And also an illustration that Σ(q∙Δp) > 0 might not be enough.

@Conrad,

Try this MAC AAPL-TLT strategy which 1.5 times outperformed Buy and Hold AAPL.

# -----------------------------------------------------------------------  
stock, bond, ma_f, ma_s, lev = symbol('AAPL'), symbol('TLT'), 55, 56, 1.0  
# -----------------------------------------------------------------------  
def initialize(context):  
    set_benchmark(stock)  
    schedule_function(trade,date_rules.every_day(),time_rules.market_open(minutes = 1)) 

def trade(context,data):  
    if get_open_orders(): return  
    mavg_f = data.history(stock,'price',ma_f,'1d').mean()  
    mavg_s = data.history(stock,'price',ma_s,'1d').mean()

    if all(data.can_trade([stock, bond])):  
        if mavg_f > mavg_s:  
            order_target_percent(stock, lev)  
            order_target_percent(bond,  0.0)  
        elif mavg_f < mavg_s:  
            order_target_percent(stock, 0.0)  
            order_target_percent(bond,  lev)

def before_trading_start(context, data):  
    record(Leverage = context.account.leverage)  
'''
Total Returns  
16651.7%  
Benchmark Returns  
10801.48%  
Alpha  
0.17  
Beta  
0.53  
Sharpe  
1.34  
Sortino  
2.10  
Volatility  
0.27  
Max Drawdown  
-35.02%

10000

START  
05/01/2002  
END  
12/29/2017

'''

@Guy

As always, I left some space for yours incredible ability to increase productivity 10x just changing one number.

@Vladimir, if I was dealing with just AAPL as in this scenario, I might just go, with hindsight, for some leverage, and see the strategy literally fly.

I took @Dan's trading interval, and restructured AAPL and TLT's participation.

Not one, but 3 numbers. The total outcome: $ 1.7 billion with higher volatility and drawdown evidently. Note that the leveraging expenses would have been more than paid for with that kind of performance.

The presented equation said: Σ(q∙Δp) > B&H = q_0∙[p(T) – p_0]. You need the outcome of your trades to exceed the profit that could have been had holding the stock. There are a lot of trade slicing combinations that could do that, just as there are a lot that would not succeed.

Here, I simply applied some leverage: 1.6∙Σ(q∙Δp).

@Vladimir, there is also that if you wanted more, you could do even more of the same.

This exercise shows that it could have been done, had someone had the determination and conviction to execute such a plan. AAPL has been doing good for many many years. Besides, we all knew that it was to the present day. And the bet should be that AAPL will continue to do so.

Using leverage in this case would have been this show of determination, of conviction in what we all think was a stock that had good future prospects and should have seen its price rise over the trading interval.

Just in case you had even more conviction. Again, the same 3 numbers were changed.
Outcome: $ 4.4 billion. Leveraging expenses, almost trivial considering, but still substantial.

Note that instead of using an in or out scenario, I used a partial out with a partial move to bonds. The reason is simple. You have a 55.5 day moving average that has a lag of 27.75 days. Like saying you are quite late to the party. Nonetheless, it does provide a “regime” switch even if it is not that good.

@Dan showed that there was nothing in the original script, and he is right. Its regime switch had no value, in the sense that it did not outperform a buy & hold.

Now, with the partial switching, some of the AAPL shares will be sold after a top, and repurchase after a bottom. Sufficiently to pump and dump part of the inventory at every MA crossing. The result, the bet size, on average, will be increasing. Also, most of the leveraging will be on what is declared as the rising part of the price movement.

Since the MA crossing was often wrong, you still continued to benefit from not having sold all your AAPL shares. Thus, the excess return. Nothing seems to beats the big bet on the big thing. But even here, the big bet (putting it all on one stock) appears reasonable, and might even be considered acceptable.

@Vladimir, hope my last post was sufficient to fulfill your request:

“As always, I left some space for yours incredible ability to increase
productivity 10x just changing one number.”

You must have noticed the change in strategy behavior. It is not due to the program's logic since it did not change. As a matter of fact, none of the trading points have been changed. Trades are initiated at exactly the same times as @Dan's script. It is a trading strategy that could be done by hand. It certainly is not complicated.

The change in performance is due to putting the pressure where it can count, on the bet size. See its payoff matrix: Σ(H∙ΔP).

Those 3 numbers changed the nature of the program from a run to safety on any sign of price weakness proposition to a trading strategy with a core position over which some trading takes place.

The 55 and 56 SMAs smooth out the data, but at a cost of being late by some 5+ weeks. Nonetheless, it provided an in and out signal, even if it was wrong a lot of times, and having whipsaw issues in consolidation areas.

The strategy's logic flow resembles the following chart which has been posted here before:

Instead of totally switching to bonds, part of the AAPL inventory is sold after a top (at the SMA crossing), it is done by reducing the leverage. From there, switching part of the portfolio to bonds even if the move is not that rewarding.

The SMA's crossover declares the uptrend where bonds are abandoned to put more pressure on the upswing. If there is a place to put some temporary leverage on, it is at the bottom near the start of a potential upswing.

All done with 3 numbers !

My point being, a lot of trading strategies on Quantopian operate on the on and off switch when maybe trading over a core position might produce better results.

We all have to explore to then settle with what we find acceptable.

We design trading strategies and operate them as if their logic was set in stone. Those 3 numbers showed that changing a strategy's behavior over time can also be quite productive. Even more productive than changing the strategy's trading logic.

There is a lot more finesse that could be added to this script. But, I would prefer a total redesign to gain control over those numbers, and provide the strategy with more refined objective functions.

Hope it helped in some way.

@Guy,

You not only met my expectations but exceeded them.

BTW Conrad and Dan started their back tests from 2002-03-15.
The early date you may start yours 2002-03-22
I run it from the date.
From tear sheet below I see:

Annualized Specific Return 45.36%

As I understand this new artificial metric it is a good number.

But there are Risk Factors which need attention:

technology 1.82
momentum 1.30
size 5.30
value -1.13

Do you have any idea how to neutralize them without loosing the only real metric - return?
Do you think they need to be neutralized?

How did you call "ma_crossover_handling" in rebalance before ever defining it??

@Vladimir, sorry about the start date mishap, being dyslexic, I often do errors like that.

But, it does not change the strategy. Had @Conrad and @Dan used the same start date, we would all have had the same SMA 55-56 crossings and therefore, the same decision points.

You raise interesting questions. We started with a trading strategy that makes some profits because at times it is in the market in a stock that is generally rising. So, we should have no surprise in the fact that it did make some money. @Dan showed that performance is also something relative, and that, whatever trade slicing used, a strategy should at least outperform a simple buy & hold.

Since I was using the same decision points. I gamed the strategy. I rigged its on-off betting system so that from a switcheroo strategy it morphed into a core position (APPL) over which on-off excess leveraging was performed.

Here are the important points from your tear sheet.

Yes, CAGR through the roof. At a cost.

Drawdown: 55.8%. Slightly more than the SPY over the same period. But, not much. Mr. Buffett had slightly over 50% drawdown too over the same period. Notice that APPL is about 20% of QQQ, it could be classified as a bellwether stock.

Gross Leverage: 1.59. Was allowed 1.0 to 1.9. Averaging out to 1.59. Note from the Exposure chart that the leveraging is really on and off based on the SMA crossing which would be the same for anyone using the same lookback periods.

You can go for leverage only if your trading strategy can pay for it and then some in order to show a net profit from using it.

Portfolio Allocation Over Time does show how much of a switcheroo strategy we have. But, this serves as the strategy's protective measure. You can accept to put leverage on rising prices, but not on declining prices where it would hurt your portfolio. It is also why you can add some leverage, if it is to be applied at the start of a potential upswing, even if at times you get it wrong.

Daily Turnover, in magnitude, is being reduced over the years. The reason is simple, your core position is getting bigger and bigger. As a consequence, the daily turnover tends to have less impact.

Daily Trading Volume shows that with time, the number of shares traded does increase substantially. Again, the reason is simple. The switcheroo thingy forces to almost double-up its core position which is growing to ever higher levels.

Transaction Time Distribution is a deception. All trading orders are sent at 9h31. And it takes all day for trades to fill. And based on the look of it. It does not succeed all the time, meaning that quite a proportion is rejected. Notice that it has the same kind of distribution as what we see as daily trade distributions for hyperactive stocks. The reason it takes all day, and that at every minute there could be some trades is mainly due to TLT. 30% of the core equity is too much to handle under the 2.5% of volume rule.

However, you do not really need TLT to trade at all. Its only purpose might have been to provide the mechanics for a switcheroo haven. But, that is handled by the SMA crossing. As a matter of fact, if you had zero trades in TLT, you would have about the same performance level, if not better. Put the downside AAPL leverage at 1.0, and TLT at 0.0, and it will add $60 million in profits.

Cumulative Common Sector Returns Attribution chart shows that all is in Technology. Evidently. We are also mostly playing momentum. What else should we expect? This betting system trades one stock and will accentuate its beta and volatility on its upswings.

To answer your last two questions. To partially reduce the impact of trading a single stock, you simple need to add other stocks. But, this might reduce overall return. As you try to average things out, that it be by diversification, or out of sync momentums, you will be dampening volatility.

This strategy wins because we did put emphasis on its volatility and did put more pressure on its upside volatility moves. Pumping at each step for bigger and bigger bets while switching all the time to safer grounds at the slightest sign of a downside price move.

We are in it for the money, and it is a CAGR game, where time, and return can have a major impact. The questions should be: what is the difference between: 1M∙(1+0.10)^20, 1M∙(1+0.50)^20 and 1M∙(1+0.60)^20? And, how much volatility can we support?

There are many ways to reduce portfolio volatility. Q is all about reducing it. But the money is made embracing volatility since by definition we need: ΔP > 0 to make a profit. The how we get it is just a matter of choice and determination.

Trying to find ways for ΔP to escape its short-term zero tendency: E[Δp] → 0, is what this game is all about.

Can these techniques be applied at a diversified portfolio level? The answer is yes. Should you want to investigate further, see for instance:

What Is In Your Stock Trading Strategy?

Trading a Buy & Hold Strategy. A Game You Can Play