Calendar Backtesting the COVID-19 Crash (Part 7)

I have been discussing impressions from calendar trade backtesting the 2020 COVID-19 crash.

One loose end from Part 6 was directional bias in my second backtest (top graph):

I calculated strike price minus SPX price (“Diff”) for 46 cases of trade entry or adjustments. In the top row, positive (negative) numbers indicate a bullish (bearish) bias. 3.41 points represents a mean bullish bias of 0.1% based on the mean SPX price 3654: hardly anything at all. Standard deviation for positive/negative differences ranges from 7-12 points, which suggests 3.41 will not be statistically significant.

My backtesting guidelines for the bullish single calendar are as follows:

• Enter single PCal in monthly expiry with short leg at least 56 DTE and delta -0.60 to -0.65.
• Exit no later than 21 DTE.
• No additional downside management.
• Upside management: if up at least 20% and underlying crosses above strike, exit trade and start anew.
• Upside management: if underlying is above strike and TD falls below 4, exit trade and start anew.
• Assess $0.21/contract for transaction fees.

Here are the ONE results:

Profit factor is 3.9 with 88% winning trades. This is really too good to be true, and you know what that means…

Directional bias here is ridiculous. The market won’t always be rallying like it did from Mar 2020 through Oct 2021:

Mean bias is 3% bullish with the average DIT 32. This implies an average annual SPX return of 34% when over decades this is said to be closer to 9-10%. For the longer term, I can see leaning 1% bullish. Perhaps that is my next backtest to run.

This strategy is dangerous because it lacks downside management. Trades beginning on 7/26/19 and 2/21/20 have maximum drawdowns of 90% and 84%, respectively. Being so close to -100%, this does not leave much room for a stop and it can destroy account equity (based on three losing trades, which is a very small sample size). The risk graph shows when leaning this bullish, losses can be hit even with the market slightly down:

I’m also including this graph to get additional perspective on why DCs are not as advertised. Note how this tent peaks around +191%. With a DC in place, even if one of the two calendars suffers huge losses (e.g. 50-100%), the potential for +191% is enough to offset that. Unfortunately, +191% is not available 32 days into the trade. At that point the highlighted line, which peaks at +27%, is a more realistic estimate (assuming no IV changes). If one calendar is up 27% and the other is down 54% then we will still have major losses to deal with. Holding until 14 DTE gets the highlighted line closer to +68%, which still falls short if the second calendar is facing drawdowns akin to those mentioned two paragraphs above.

As option traders, we would probably benefit by hiding the expiration curve entirely since it’s usually too good to be true.

Calendar Backtesting the COVID-19 Crash (Part 6)

Today I continue discussing my impressions from calendar trade backtesting the 2020 COVID-19 crash.

I left off showing results for a single-only calendar strategy, which is a contrast to guidelines from Part 1. When I said these results are not encouraging, I was comparing with backtest #2 (summarized in Part 3) whose ONE results look like this:

PF is 2.2 and average win is 2.5x average loss. Net profit is much better. I would certainly consider trading this. The DC allows me to hold a losing calendar longer because it also includes one that is winning.

I’m still not totally sold on DCs. Perhaps single calendar results can be improved if I hold a losing trade longer.

This is really suggesting a directional bet. Holding a losing calendar longer means dropping the TD > 4 criterion. This usually happens when the market is above or below the calendar strike, which implies a bearish or bullish bet, respectively.

Given directional choice, I would certainly bet bullish. I mentioned the upward bias to equities in the fourth paragraph here, in the last paragraph here, and in the tenth paragraph here.

Before I decide how to implement this, I want to consider how even the ONE results shown above might be improved. Number of losers still exceeds number of winners, which is not ideal. Also, profitability is concentrated between Apr 2020 through Feb 2021 (10 months). The rest is equity curve chop, which makes me question consistency. A larger sample size of years leading up to 2020 would broaden the context.

I should check to see what directional bias I already have in the trade. I did not intend any, but starting each trade at the nearest 25-point strike implies a tiny bit. Number of times I leaned bearish and bullish should roughly balance out.

I will continue next time with backtesting a bullish calendar trade.

Incidentally, I clicked [the lower box] to see the following ONE results:

Compare this to the graph shown above. This also has a PF 2.2 and number of winners exceeds number of losers. This appears to be a more consistent strategy.

Unfortunately, these results are a mistake. I accidentally turned off commissions, which eliminates transaction fees. This is fantasyland and we have known this for a long time (see third-to-last paragraph here and fourth-to-last paragraph here).

If I significantly overestimated slippage, then the single-only calendar strategy with no directional bias might be worth doing. This would then be another instance like that described in the second paragraph here. Any backtest where I stack the odds against and performance still comes shining undeniably through is a strategy with which I really want to go live.

Calendar Backtesting the COVID-19 Crash (Part 5)

Today I continue discussing my impressions from calendar trade backtesting the 2020 COVID-19 crash.

The bigger point with which I left off is that a cursory check (at least) to authenticate a backtest should always be done. No backtest is Truth and live trading is often not exactly per guidelines. Some traders are more discretionary than others and some couldn’t replicate a set of backtested trades even if they really wanted to. Nevertheless, in a backtest like this or this with a limited number of trades and a short time horizon, one mistake can materially affect results.

I’m not sold on double calendars (DCs) because overall, they don’t seem to grow profit as quickly as singles. This is subjective and anecdotal, but I’m going to try and put some reasoning behind it.

DCs are more risky than single calendars for two reasons. First as mentioned in Part 4, being established through adjustment implies less time to recoup transaction fees. Second, being established later means more gamma risk where substantial market moves can result in big losses. DCs need not be riskier from a margin standpoint, but the greater gamma risk is pervasive and insidious. This matters in the face of large market moves, which are a reality in trading.

To control gamma risk, I could implement a minimum DTE limit beyond which only closing contracts and re-entering farther out in time would be permitted. Theoretical determination of this optimal drop-dead date would be very challenging.

Another possibility would be to allow just single calendars. When an adjustment point is hit, exit and enter anew with at least 56 DTE. Here are the OptionNet Explorer (ONE) results of such a single calendar backtest:

This is not encouraging. Profit factor (PF) is only 1.2 and number of losers exceed number of winners.

In running this backtest, I felt the market often moves around too much for a single calendar to contain it. This led to more frequent whipsaws than backtest #2 (Part 3) where DCs are permitted.

In running this backtest, I also became very cognizant of how quickly a single can lose money when the market moves away from the strike. DCs looks appealing with the wide profit tent, but one calendar is always losing money as the market trends. This may explain why profit seems to develop slowly. Day steps would do a better job at revealing extent to which the expiration curve is mere mirage. This is usually the case with my option trading, but perhaps it’s more pronounced here.

I will continue next time.

Calendar Backtesting the COVID-19 Crash (Part 4)

I left off asking about a subtlety that would probably be identified by only serious option traders. Submitted for your approval is a question about why one backtested calendar is ~30% more expensive than all the others. Were you able to explain it?

The nuance here is that calendars generally get more expensive when purchased closer to expiration. To lower the cost, I would generally go farther from expiration, but that is not the case with the Sep/Oct calendar I mentioned in the second-to-last paragraph of Part 3. The calendar I backtested was an Oct 3225 put priced at 24.09% IV and a Nov 3225 put priced at 26.24% IV. Is that normal contango? I’d have to survey a large sample size to find out but my guess is no.

I suspect two factors have contributed to make this calendar atypically expensive. First, it is 35 days wide. This happens four times per year [(4 * 5 weeks in between mos) + (8 * 4 weeks in between mos) = 32 + 20 = 52 weeks for the full 12 mos] and it means paying for one extra week of extrinsic value in the long leg.

Second, and what I believe to be of greater importance, is the 2020 presidential election. I think the Nov options are priced rich due to election uncertainty.* On 7/31/20, IV for Nov and Dec are 24.55% and 24.66%, respectively (beyond which term structure decreases). Purchasing the Nov/Dec 2020 112/140 (DTE) calendar would cost $4,650: much less than the $8,840 seen here. The horizontal skew on the Nov/Dec would be -0.19, which is much lower in magnitude than -2.15 for the backtested Oct/Nov. Term structure rises steeply through November and then starts to trail off. This is unusual.

Incidentally, I’m puzzled why I exited this adjusted DC at 43 DTE when a third adjustment point was hit on 9/3/20 to open a single Nov (78 DTE)/Dec calendar. I could have rolled to ATM, as the presenter’s guidelines suggest (green):

I could have also gone an extra 100 points OTM (green):

Theta is comparable for both and cost is slightly less for the latter. TD is even higher for the latter (17 vs. 7). Either of them look doable. Every time I adjust a calendar, PnL gets dragged down a bit by transaction fees. Since it usually takes a few days of positive theta to recoup that, maybe I should never adjust too close to my drop-dead 21 DTE. That is no excuse here with 43 DTE and plenty of time left.

I will continue next time.

*—As an exercise for another post, I should plot this term structure in Python. An additional
     detail would be to scale the x-axis properly since data points start one month apart and
     later increase to points that are three months apart.