Exploring Kaeppel’s Sector Seasonality (Uptrend Confirmation)
I’m poking around the edges of Kaeppel’s Sector Seasonality Strategy. A refresher on the strategy’s performance from 1987 (in red):
[logarithmically-scaled, growth of $10,000]
Kaeppel’s original strategy has been prone to some pretty significant downturns (ex. year 2000), so in this post I’m testing only taking the month’s proposed trade when that sector is in an uptrend (otherwise remaining in cash for the month).
I’ve defined being in an uptrend as the sector’s 50-day simple moving average (SMA) being above its 200-day SMA, but note that I’ve tested other similar measures with more or less the same results.
New (green) vs original (red) strategy results from 1988:
[logarithmically-scaled, growth of $10,000]
In terms of just return, requiring uptrend confirmation has clearly hurt the strategy, but that’s only half the story. The numbers below show some simple risk-adjusted metrics. For comparison’s sake, I’ve also included the inverse of our uptrend confirmation rule (i.e. requiring downtrend confirmation to take the trade).
Geek note: these results are from 1988 (because nearly a year is required for the 200-day SMA), so they will differ slightly from our original results from 1987.
In terms of risk-adjusted performance (Sharpe and return versus drawdown), requiring uptrend confirmation hasn’t helped, and might have actually hurt a bit.
More Likely to be Curve-Fit?
As we wrote when we first introduced this strategy, it is by its nature very prone to curve-fitting (meaning that the likelihood of it working in the future might be lower than normally expected).
It would seem to me that if 50/200-day SMA crossovers, an approach to long-term trading that has proven historically to be a pretty robust indicator for reducing drawdowns and downside volatility, failed to do just that on this particular long-term strategy, it might be a pretty strong clue that this strategy is particularly curve-fit.
That’s a bit of “thinker”, but mull that over for a moment. Not only is this strategy purportedly picking bullish months for these particular sectors, but it’s picking bullish months that more often than not buck the prevailing trend. That’s a pretty tall conclusion for a strategy with so few trades to analyze; not a death knell mind you, but something to think about.
In a follow up post we’ll break down individual sector performance. More to follow.
[click for a summary of all recent posts about Kaeppel's strategy]
Happy Trading,
ms
. . . . .
To stay up to date with what’s happening at the MarketSci Blog, we recommend subscribing to our RSS Feed or Email Feed.
Filed under: Stock Market Sectors, Time-based, Trading Strategies | 6 Comments
I did some less-than-perfect analysis that was interesting to me. From 2000, using monthly beginnign trade prices and 10 month moving averages.
Equity returns improved if the 1st trade day was above 10 month moving average. Leverage also improved returns. I avoided equities if below 10mma on beginning trade day.
Energy was the most consistant. 10 mma did not effect returns. In fact it gave greater returns if the trade day was below the 10mma. My rule for energy evolved into: use it.
Gold producers was similar to Equity: above 10 month moving average improved returns. Below 10mma: cash.
Was reviewing the seasonal sector strategy. I’m somewhat torn in the analysis.
I would be the 1st to scrap the system if I suspected a curve fit scenario (Hey, all make believe…Yay) or if all of this was well, just coincident. The fact that I know seasonality, especially in NRG, is as real entity. Whether it’s a run up in WTI front month in the spring (evidenced here) or a move in the spreads from inverted to contango every fall, & those (3) charts really have my attention. Albeit, the gold thing has me wondering.
1) Average Monthly Return by Subperiod – 1995 – 2002 seem to be more of the anomaly. Yet, 1987 -1994 & 2003 – 2009 seem to be more consistent in return.
15 years of 1.85% vs just 7 years of 3.2%
2)Average monthly return by year – The same statement seems to hold true with
1998 & 1999 seem once again to be anomaly. Yet, the others years in the sample seem consistent in to a range.
3)Sector Seasonality Return minus VFINX Return by month 87 -09 – exhibiting a trend line upward slope strengthening & of course.
“This persistence argues against both data snooping bias as an explanation of past performance and market adaptation after publication.”
How can we verify or disprove the presence of curve fit to truly “kick/ shoot holes in the tires” here. As on preliminary bite down on this nickel one is tasting more metal than wood.
Or is it me? As I’m a newbie to this data analysis stuff. Where are my holes?
Best,
JP
PS…Also, do u have a link explaining your sharpe & vol math/numbers?
RE to JP: great comments sir. Some thoughts:
I think it’s difficult to capture mathematically whether something is curve-fit (except w/ out of sample results). I think the best measure of curve-fit is logic, sample size, sample length, and sample consistency. Sample size and length are easy enough, more trades that consistently perform over a longer period of time are an indication of robustness. Sample consistency would mean that gains are not coming from a handful of good trades, but are being contributed to by the whole sample.
The logic criteria is more subjective. In this strategy, the author basically said “let me find the sector that’s performed best in January…next, let me find the one that’s performed best in February…etc, etc.” Especially given the small sample size, that approach is data mining is by its nature going to produce curve-fit results.
Note that I don’t think that means the strategy doesn’t work (I think there’s some truth in these seasonality plays) but I do think it means that future returns aren’t going to be as stellar as this backtest.
P.S. no good links on this blog re: Sharpe Ratio and volatility numbers, but both are common metrics. Some google searches should turn up some good info pretty quickly.
Thanks for the thoughts.
michael
hi michael, nice work……my comment would be that the transitions based on seasonal cyclicality may not be captured by such a long term indicator like the golden cross. I think confirmation may best be captured using the 10/20 ema crossover which I would test myself if i had the Fidelity fund data personally.
best
dv
RE to DV: hello good sir. Good comment.
Two thoughts: (a) because this is a monthly system (only trading on the last day of each month) I can’t use an indicator that’s too short, but (b) having said that, I tested other shorter MA combos like 10/20 as well with more or less the same conclusion. I stuck with the 50/200d for simplicity’s sake.
michael
Wonderful post! Great examples of both testing a strategy and then questioning whether it’s really not just data snooping.