A reader inquired about the “Short Term Stock Selector” designed by Robert Hesler, which “has provided neural network generated swing trading predictions [approximately daily] since 1996. …All buy and sell recommendations are based on 19 technical indicators. Some indicators pertain to the market in general while others pertain to individual stock attributes.” We focus for this review on the most profitable type of trades (Type A), for which the site claims a 66.3% win rate and a 40% annual return on investment over the period 4/11/96 through 6/17/08. Using the detailed listing of 12,340 closed Type A trades over this period, we find that:
We make the following assumptions regarding this trading data:
- The buy and sell prices represent actual, not assumed, values.
- The buy and sell prices do not include broker trading fees, so we test a range of fees and initial position sizes per trade.
Overall trading metrics are:
- The average win rate for the 12,340 trades is 66.3%, based on gross returns (excluding trading fees).
- The average gross return per trade is 0.96%, with standard deviation 8.23%.
- The average holding period is 10.3 calendar days (7.1 trading days), with standard deviation 7.2 calendar days.
- The average number of positions open at any time is about 29.
How do trading fees impact profitability?
The following chart summarizes the net return per trade (including trading fees) for various combinations of round-trip trading fees and initial position sizes. Results show that small initial positions sizes may be unprofitable on average and that bigger initial position sizes are better.
Is profitability consistent from year to year?
The next chart depicts the average gross return per closed trade by calendar year. 1996 and 2008 are partial years. The chart shows that average gross returns vary, but only one year (2001) has a negative average gross return. A best-fit line suggests that gross profitability of the system deteriorates across the sample period, but the small annual sample size provides relatively low confidence for this finding. It is arguable that the first full year in 1997 represents a lucky start.
Does the combination of trading frequency and average trade profitability suggest good macro timing?
The next chart adds the number of trades per year to the preceding chart (excluding partial years 1996 and 2008). If the system times effectively at a macro level, trading frequency should be high (low) when the average gross return per trade is high (low). Although the annual sample is small, results do not indicate that the system times trading effectively on a yearly basis. For example, trading activity is higher for a poor-return 2001 than for the above average-return 1999 and 2000.
Is the trading system practical from a capital requirements perspective?
As noted above, the average number of open positions is about 29, and small positions eliminate or diminish profitability due to the impact of trading fees. The capital required to implement the system faithfully and profitably is therefore substantial. In fact, the number of open positions varies considerably, as indicated by the following examples:
- On 10/28/97, there are 170 new positions with 59 older positions still open (229 total open positions).
- On 5/11/04, there are 155 new positions with 116 older positions still open (271 total open positions).
- On 2/28/07, there are 149 new positions with 4 older positions still open (153 total open positions).
At $10,000 per initial position, these situations require millions of dollars of (unleveraged) capital. Moreover, this clustering of positions within short intervals raises the possibility that trades unexploitable due to limited capital distort the overall average gross return per trade. To test for such distortion, we eliminate trade clustering by arbitrarily selecting every 100th trade (with the entire dataset sorted on the purchase date) and further eliminate duplicate dates within this reduced sample. The resulting winnowed sample consists of 123 trades with average gross return per trade of 0.45%, less than half the 0.96% found for the entire dataset. In other words, it appears that trading system profits do “cluster” during intervals with many open positions and that a capital-constrained trader could not achieve the higher 0.96% gross return per trade.
These spikes in capital requirements bring into question our assumption that the dataset buy and sell prices represent actual and not assumed values.
Might there be unrealized underperformance associated with this trading system?
The final chart is a plot of gross trade return versus holding period in trading days for the entire dataset. The vertical axis is truncated, excluding two instances of extremely high returns, to enhance visualization. The best-fit line through the distribution, which includes the two unseen points, indicates a clear tendency to hold losing trades much longer than winning trades. Such a tendency would normally result in bow wave of unprofitable open positions, not included in the return on investment calculations for closed positions.
In summary, trading fees, trader capital constraints and inclusion of open positions may substantially reduce or eliminate “Short Term Stock Selector” profitability as summarized by the offeror.
