I’ve never been a discretionary trader, but I have spent the last 10 years doing quant work: modeling, information extraction, and automation. I know the areas where quantitative methods are weakest . It seems sensible for discretionary traders to focus on these areas instead of struggling in areas where quantitative methods are well-suited:
1. Don’t Use Price Data
Don’t focus on price-based signals. I’m not a believer in chart reading or technical analysis. Humans are well-known for our ability to see patterns where there are none. Price (and volume) data is precisely the area where all quants focus, so it is too competitive. Every computer scientist who gets into machine learning immediately tries it out on stock prices, because the data is readily available and requires no domain knowledge. Don’t compete in this area without the cutting-edge tools quants use.
2. Small Sample Sizes
Quants don’t trust models trained on few examples. The risk of a false discovery is too high. So, that is an area where money might be left on the table. You can counter the risk of a false discovery with domain knowledge. The more expertise you have, the fewer historical examples you’ll need for confirmation. Stay away from trading strategies with plenty of examples, as they will be too competitive. Despite all the hype, techniques like deep learning don’t work well without a lot of training samples.
3. Unreliable Backtests
Focus on strategies where backtesting is unreliable. I’ve recently been collaborating with Glenn Chan , who saw my post mentioningEDGAR filings. He digs through regulatory filings looking for red flags which might indicate shorting opportunities in smallcap stocks. He’s had very good luck, however backtesting his ideas is of limited use. In some cases, his picks had no shares available to borrow. Others were so popular among short sellers that borrowing costs were higher than credit card rates. I don’t have historical hard-to-borrow data, so my backtests can’t really be trusted. In general, short strategies in smaller stocks tend to be avoided by quants, which makes them good for discretionary traders.
4. Overnight News
Trade only on news that comes out when markets aren’t open. You have a massive disadvantage on speed. Reading overnight news gives you hours to digest it at a deeper level than quants can, at a time when their speed will give them no advantage.
5. Labels and Features
Machine learning, both supervised and unsupervised, requires features. Focus on datasets with features that are difficult to encode. Take for example, a model estimating the worth of a house for sale. The usual features include square feet, number of bedrooms, etc. Your model should include additional features that others don’t have, like how well the neighbors keep up their yards. This information is not easy to collect, which is exactly what makes it valuable.
Supervised machine learning requires labels in addition to features. Look for situations where labels are hard to come by. For example, I once tried to systematically analyze Schedule 13D filings . As shareholder activists build up a big position in a company, they have to disclose it in 13D filings. They typically say they are buying the shares “for investment purposes only.” Then after a while, they file an amendment that declares their true intention of shaking up the company. The problem I found was that the investment-only filings used the very same legal terms and language that the important amended filings used. They would list all the things that were NOT their intention. Document classification in NLP is often done using a bag-of-words approach for features, which turns out to be useless for this situation. Getting labels was too hard. Quants don’t generally read and label 50,000 legal documents. The language was too dense to crowd-source to the Mechanical Turk . Since I didn’t have labels, I moved on to other ideas. (Although, I’ve since learned some techniques that make me want to revisit this idea!)
Statistical models often include shrinkage factors or regularization terms. Such tricks keep the model from being overly influenced by outliers. Since quants are actively trying to ignore outliers, you might have luck actively seeking them out and analyzing them.
7. Quality vs. Quantity
Following quality recommendations may be better than following the aggregate voice of the masses. Twitter is full of stock picks that are likely uninformative. I once had access to a full database of analyst recommendations. As a quant, I didn’t get to know the individual analysts or the quality of their recommendations. I had too few samples per analyst to see if the data could identify the skilled ones. So, I went the easy route of treating all analyst opinions equally. As a discretionary trader, you are positioned to better understand which sources of recommendations can be trusted.
8. Illiquid Stocks
Big funds have no interest in smallcap stocks. There isn’t enough capacity to make it worth their time. This makes them less efficient, and a better playground for discretionary traders. Furthermore, quants use transaction cost models to estimate impact. These models are less reliable for small stocks. Even the price data is less reliable for small stocks. For example, it could be the last trade of the day occurred 15 minutes before the close. So a model based on daily data might find an inefficiency that wasn’t real. See #3 about unreliable backtests.
Stay away from machine-readable news. Actually, the machines are steadily encroaching even on the non-machine-readable news. My example is the Form-4 filing. Starting in about 2003, these filings have been presented to quants in a very convenient XML format, like this . Before then, they were much harder to parse. The market reaction to these filings is now instantaneous. More and more filings types are becoming machine-readable, which should make them less profitable to discretionary traders.
10. Messy Situations
Get involved in messy situations. Quants don’t. I had a friend who got rich buying junk debt for his firm. He untangled legal messes to evaluate their chances of getting paid. For example, when Bernie Madoff’s pyramid scheme came crashing down, investors lost a ton of money. But they didn’t lose all their money. Some of it would be repaid after the government had time to sell assets and sort out who gets what. My friend bought the rights to some payments from Madoff investors who just wanted to get what money they could right away.
转载本站任何文章请注明：转载至神刀安全网，谢谢神刀安全网 » 10 Tips to Help Discretionary Traders Compete with Quants