Sunday, September 06, 2009

Cross-Talk: Automated Systems and Discretionary Trading

In a recent post, Henry Carstens has offered two rules of automated system development. He suggests that the odds of creating a viable system vary directly with the system's holding time, and he conjectures that systems with shorter holding periods will be more likely to degrade in performance.

Henry has more experience in system development than I ever will have, but my limited experience supports his hypotheses. I have found that the largest "edges" in trading occur over holding periods measured in days, not minutes. I have also seen the greatest fall off in performance among discretionary traders whose holding periods are measured in minutes, not hours or days.

My conjecture is that the technology arms race that has led to automated market making and increased high frequency trading has created a high proportion of noise around any possible very short-term market signals. Though markets may still move from Point A to Point B over a period of hours or days, that automated trade has greatly influenced the path from A to B.

I recall, in my earlier days, that experienced traders used to teach beginners who were learning to scalp to limit losses to one tick losers. Today that advice would be silly. It's not unusual to see buy or sell programs come into the market and take out several ticks at a time.

Over longer holding periods, that noise is easier to sit through: a "swipe" in the market that takes out three ticks is not a threat to a swing position targeting a move of 10 ES points; it is a potentially fatal threat to a scalper hoping to risk a few ticks to make a point or two.

Henry's conclusion is that, on average, automated trading systems should outperform discretionary trading over longer holding periods. Discretionary traders with a feel for order flow and momentum can adjust to shifts in participation close to the markets in a way that systems cannot. Systems, on the other hand, can sit through noise and capture signal in a way that often eludes traders beset by perceptual and cognitive biases.

Might it be the case that valid, backtested systems offer an unusually promising platform for teaching new traders how to trade? The systems provide setups with an edge, requiring traders to learn the execution and trade management skills to make the most of the trade. The degree to which traders outperform the system signals reflects the value that their discretionary trading brings to their accounts.

Perhaps one definition of a good discretionary trader is one who adds value to an established trading system.


Jorge said...

Dr. Steenbarger,

Yes, one of the problems with medium-term automated systems is that it's really hard to program some decent timing into a system without falling into over-fitting. That's where a discretionary trader can add (or subtract) value.

The problem is that many traders confuse trying to get a better entry price (basically trying to minimize slippage in trend-following and breakout systems and/or avoiding price shocks by not getting in right in front of a major news announcement) with cherry picking entry signals (which defeats the purpose of having a system in the first place).

Best trading,


SerpentMage said...

Interesting that you write this article. I recently wrote a trading system that had stop losses and while it trained it did well, but in real life the system just collapsed. It underperformed.

Though a few years ago I trained a system that had no concept of stop losses. It actually filtered out the noise. That system made quite a bit a of money. EVEN when running life.

My point is that yes there is a signal to noise ratio and it is getting worse not better.

2positive said...

Hi, Brett, I'm a long time reader of your blog and I love your books too.

I've been developing trading systems for about a year now and came to the same conclusions about time horizon. My trades last from 2 days to 2-3 months.

Another hugely important conclusion is that stoplosses are actually evil and hurt long term performance of almost any system. They are meant to control risk, but they actually increase it. There's allways a better way to protect you from big move against you (diversyfying, limiting your maximum loss with options, etc.) and a better way to exit a trade then using a % stop loss.

So now I trade almost exclusively with option vertical spreads on stocks.

Sabretache said...

Thoughtful, interesting stuff as always.

As someone now in his seventh year trading a small personal retail account full-time, permit me a few observations.

Firstly, by way of pedigree: In my first year I lost @ $60K, I had finally clawed that back by the start of year 5 and have now traded profitably for 2 years or so. However, profitability such as it is, is causing me to question the viability of the entire effort going forward. The reward * risk / time commitment involved is probably worse than any other undertaking I have been involved in over a 40 year working life to date. Discount the risk element (a pretty stupid thing to do) and the whole thing is still marginal at best. So to my observations:

I started with purely discretionary, swing trading with holding times averaging maybe 2-5 days. I gradually moved to shorter holding times and slowly developed automated systems using Ward Systems Neuroshell suite of products. Last year (Just $25K profit before subscription and other expenses), commissions were just over $29K - (The brokers just lurve.... high-frquancy traders eh?).

Which brings me to my main point. It is that the problems associated with constructing profitable high-frequency automated trading systems have probably got more to do with the proportion of the average trade profitability accounted for by brokerage plus the spread forced on all retail traders, than the unpredictability of so-called 'Noise'. And for ultra-high frequency (ie scalp-type trading) you can add the extent to which every one of your trades itself becomes the potential precipitator of the sort on 'mini-move' that no amount of AI analysis and back-testing can genuinely account for - ie the trading equivalent of the Heisenberg uncertainty principle.

Personally I find the concept of 'noise' a difficult one, to the extent that it is probably only valid in a relativistic sense - ie relative to the per-trade swings acceptable to the particular system defining it. In my experience the upper and lower limits of tick level 'noise' is largely defined and controlled by mareket-making activities - ie by the few traders (often behaving to some extent collusively) who dominate the order book of the instrument traded.

PS - The above observations apply strictly to medium-high liquidity stock index futures contracts though, with obvious limits, they probably apply to other markets to - commodities and pm's for example.

The Tranquil Trader said...

This is probably the article i hold most resonance with. Any short term systems (in mins) will tend to get eaten alive by the bid/offer, slippages and data accuracy.

mixing a several systems (with different underlying concepts) tend to offset drawdowns whilst keeping returns intact.

I have found putting stops tend to destory returns too because of the mean reverting nature of the markets i trade. But instead but i do put on BIG stops (a good 2x ATR or more) just to avoid the one off those long tails events).

Just wonder what the consensus is when it comes to applying stops on systems..

D TradeIdeas said...

Henry Carstens' 2 Rules of Automated System Development hold true in my experience building odds-based trading plans within Trade Ideas.

Rule 2 makes sense: the market changes more frequently and radically in the short term - often invalidating certain themes that short term strategies are built on. The skill in Rule 2 becomes knowing when such themes and strategies the market is steering towards and which ones from which its moving away.

Rule 1: Approximately 3 out of 4 of the strategies I backtest benefit from the longer hold times. This is especially true if such strategies begin trading after the first 30 minutes of the market open when frenetic, hectic activity is at its peak.

We do find highly successful strategies during the opening 30 minutes, but the holding times for these trades are intraday - sometime only for as long as 15 minutes. The automation of strategies during the opening 30 minute period is the only method used to produce our positive results - manual trading is too slow (many autogenerated offers are purposely away from the market and therefore only a % gets filled).

Market participants in general and our subscribers in particular with less trading experience do better with longer holding times as the dynamics of the trade can be digested and the market volatility/programs better understood with a longer (1-2 day) horizon.

Indeed our most advanced users of our technology are the most adept at using the same Idea Generation and Event-based Backtesting tools of Trade Ideas to execute a scalper's trade.

Great article.