The QuantBeat
ABCs of Alpha, Beta and Consistency - A QW Solution
by Sunil K. Pai on 06/07/11
Ok, so where is our sustained alpha? We attribute it to the following:
- Style and factor adaptability: humans (fundamental analysts) might focus on value (mean reversion) or growth (momentum) and have periods of outperformance but invariably will run into periods of being out of favor; we have no such bias in our investment process so our returns have as much opportunity to capture mean reversion as momentum
- Multi- and high frequency rebalancing: In today's high-frequency, highly liquid world, markets are as efficient as ever and unless one is capable of capturing this excess motion while managing risk, then excess returns are not likely to be sustainable
- Automatic, inbred risk allocation process: Not all investment decisions are equally advantageous and being able to align exposure to the probability of profit is integral to creating an opportunity for excess returns
- Combining both forms (reversion and momentum) of stock returns into our decision provides not only a greater potential for capturing return but also provides an offsetting risk function since reversion and momentum have negative correlation when explaining returns, which positively impacts risk adjusted returns
- Low potential for tracking error: Our model and live portfolios are easily kept in sync and rebalancing costs average about 20 to 30bps of return each year, which can be largely offset by dividend income meaning this is a practical real-world model that does not incur substantial cost to maintain
QW's Comprehensive View on Risk Management
by Sunil K. Pai on 06/07/11
There are many forms of risk that resonate within the investment management world. The ones that are nearest to the investment decision focus on exposure control, Value at Risk (VaR), volatility, correlation, position limits, capitalization, liquidity and leverage. These elements become more comprehensive when you include regulatory, reporting, administration, operational, strategy, legal and environmental. This post will focus primarily on the former to discuss how we combat the risk management quandary so that our model will be able to survive in the real world of drift, shocks and exogenous risks:
- Our trade decisioning for each position is based on 7 relatively high-frequency models, which automatically (without human emotions) reassess the exposure on a position every day. If this basic trade decisioning process will not show a profit (without stops), then there is no point to stops other than to smooth your rate of loss.
- Due to the high-frequency of model exposure reassessment, when it is wrong, it will not hang on so long as to dig a big hole on any one position. However, just because it reverses a position does not mean the model cannot be wrong many times over, but we gather from large scale testing what the max and average drawdowns are for each position. We review model outputs daily to monitor portfolio drawdowns and are happy to share this data with any interested party.
- With ETFs or even equity index futures, we have tremendous diversification of individual stock risk, so we don't worry about issue-specific risk. We definitely have to worry about market risk, model risk and operational risks.
- So, we have a minimum of 15 ETFs right now to diversify sector specific risk, each ETF model is comprised of 7 individual frequencies to diversify the model itself, and operational risks are contained by keeping trading to a minimum, data input to a minimum, # of positions traded low, transaction costs to a minimum, redundant computers, power supplies, and communications lines. All of these elements comprise the vast majority of our risk and we seek to manage them comprehensively.
- Since our model only allows one daily data point for each position to enter it, high-frequency trader (HFT) models have no chance to pull money out of our portfolio by triggering it with odd price prints, triggering stops to take our positions away before our signals run their course, playing with bid/ask spreads when we go to execute a trade, or affecting the price of our ETF (too difficult considering the hundreds of underlying stocks). Bottomline for those who want to trade in this market is beware because the manipulation is rampant in this "free market" under the guise of market-making. This manipulation element has always been there before HFTs but HFTs have elevated it to enormous proportions. And, all these big institutional mutual funds are impacted by it - many may not realize it just yet. The HFT's profits have to come from someone's pockets after all!
- Onwards to liquidity and correlation risks, those are managed by selecting ETFs that have tremendous underlying liquidity such that under even significant market duress (like 2008), there is limited to no chance for us to get locked into a position. And, similarly, we have the ability to move massive amounts of money into an ETF without impacting the price we get giving us great capacity to put AUM to work under our models. This is an inherent weakness of many quant models and certainly that of individual traders who cannot manage multiple positions simultaneously especially in volatile markets.
- All of this does not mean we have no risks, but they are different from the traditional ones. We have data sampling risks, meaning if the data for a symbol is not sufficiently representative of different types of motion, then our optimization process may not create meaningful information that is essential to our portfolio management process. This is the most key risk we have because our model "learns" what to do in the future based on what it sees in the historic data we feed it. Naturally, we use long periods of data for this learning process but the future is still an unknown. We also have to update this "learning optimization" process on a regular basis (usually once a Qtr.).
- Hopefully, you will find a resounding theme here, which is that we manage our risks comprehensively but we also don't have any pre-conceived notions about how the future should behave. So, our model does not make "predictions" though it does try to identify the potential path a stock is moving down and align the exposure of that stock according to how risky that path might be. Without predictions, we don't expose ourselves to any prediction bias (another risk) or even behavioral biases. As a result, what we do is ensure the model is properly capitalized, executed and followed as closely as possible - those are real risks to be able to survive the model when it inevitably draws down.
- From considerable testing, we have identified (by individual stock) profit and loss potential. For our current portfolio of 15 ETFs, last year's max DD was around 5% unlevered and around 1% on average.
- As a real life example of how this works, here is what happened on Monday, Feb21, 2011 as the SP500 lost 2% while our portfolio was down about 1%. We re-ran the models around 330pmEST which took us to a basically market neutral stance with some sectors long and other short versus the 51% net long position we had going into today. Days like today are never fun but the key is survival if one is on the wrong side. Note the models reduced their positions, taking PnL (either profit or losses) but strictly as a function of the model signal quality. So, let's see what happens in the coming days. We post these updates (www.quantworks.net) just before market close for those interested in seeing our model's view of the optimal risk-based portfolio composition given current pricing cross-currents.
Lastly, we have attempted to integrate the worlds of high-frequency trading with that of investing to create a high-frequency investment process, which we see as the only reliable method to generate sustainable returns with contained risks. Hope this discussion did not go too long but we like to be as transparent and informative as possible such an important topic.
May Performance Update
by Sunil K. Pai on 06/07/11
HFRX (www.HFR.com) is the well-known database for many hedge funds to report their results. The results are reported net of their fund expenses and inclusive of any leverage employed.
Hard to believe that the HFRX Equity Hedge index is -6.11% this year so far and -2.63% in May alone and this is with the SP500 up over 9% YTD May! Not good, and shaping up as a rough year for them.
Our fund-of-funds product is +11.1% (+9.3% net of 1/10% expense) YTD May and Net +1.2% for May (rough month) with no losing months this year yet (of course June could change that). The month is still early and it will be interesting to see how the portfolio lands this month. Our losing months tend to be around -1%.
Best,
The QW Team
2011 1st QTR Performance Discussion
by Sunil K. Pai on 04/11/11
All things being equal, our first quarter was quite strong and profitable
(all stats for 1Q2011):
- Booked unlevered gross +4.0%,no losing months (SPX +5.3%) on 15 ETFs
- Max unlevered portfolio drawdown of 2.5% (vs SP500 at -7%)
- Avg. portfolio net exposure <30% net long
- HFRX Equity Hedge index -3.7%
Despite the broader market gyrations off their highs during this "Black Swan" of quarters, we incurred a little pain but kept capital volatility at a low - a driving focus for our algorithms.
While we don't presently trade futures, we did test our signals against the corresponding eminis, leveraged ETFs and Single Stock Futures (SSF) contracts and obtained results consistent with the increased leverage. Execution risks are limited to none with our model as it updates once daily and has embedded risk management provisions. Our live money models are now integrating with the leveraged ETFs.
So a great start to 2011 after an excellent 2010. And, always welcome all questions or comments.
The QW Team
Not Fundamental, not Technical, What Do We Do? Part II
by Sunil K. Pai on 02/27/11
Outside of simply being lucky with one's strategy, our investment process centers first on finding a mathematically sound construct that can provide sustained alpha extraction from a liquid market. Here we look to the "phase space reconstruction skills" of our resident rocket scientist, Dr. Xi Shao. Why? Because what we have found is that by using a multi-frequency framework that encompasses time, speed, distance and volatility, we can obtain sufficient information to characterize recent stock motion into two key components -- namely, mean reversion and momentum.
Next, we need a large amount of sample data for each stock that is integral to our optimization process. The optimization process is central (and proprietary) to identifying the key mathematical factors that "explain" the returns of a stock during its in-sample period. The optimization will maximize the price-based return of the stock for each of seven unique model frequencies and allows us to frame recent motion into its constituent reversion and momentum parts. We have found sufficient stability in our optimization parameters that a year may elapse without updating our optimization and still obtain returns in the out-of-sample (akin to live) period that are consistent with the in-sample results. For those statistically inclined, this could also be read as a high t-stat and forms the basis for us knowing beforehand the potential path a stock's return structure will follow. Reality is: we re-run the optimization once each quarter for each stock in the portfolio.
Armed with in-sample based settings unique to each stock, we can then frame current, live out-of-sample data for a stock and make two key decisions: 1) do we long or short, and 2) how much to weight the position. Both decisions are automatically determined by the 7 individual models that comprise each stock. Each model systematically decides either to be 100% long or short but when added up they collectively determine our position's weighting. Since 7 is an odd number, we will always have an allocated amount of exposure to each position. For example, if 2 models are -1 and 5 are +1, then our decision would simply be Long +3 (eg. 5 long - 2 short). And, this exposure is re-assessed once each day using live price data only.
What this all means is that we carry a statistical a-priori belief that over long periods of time, our portfolio will reliably generate excess returns and a sustained Sharpe ratio that exceeds the market's and the majority of our peers, especially of the human variety. And, it bears mentioning that since we initiate both long and short exposure, our returns can be absolute rather than relative to the market performance. Our mathematics are also efficient to use for a long-only portfolio (for ERISA accounts) as our returns structure is as efficient on upward only movements of stocks as it is in identifying downward movements.
This is a lot to chew for even typical fundamental and technical investment professionals to follow (not to mention that it goes against their grain), but in its simplest form we do the following:
-
Create and optimize a portfolio of diverse stock holdings
-
Set the weighting on each stock given the highest expected risk adjusted return based on our two factor (reversion and momentum) optimization
-
Maintain a "regime" switch based on market volatility to maintain adaptability
-
Use a basket of ETFs for their low risk/cost appeal
-
Re-optimize portfolio daily (faster than any human can re-think their portfolio optimization)
-
Maintain exceptionally low rebalancing cost and insulation to HFT manipulation
-
Importantly, manage all forms of risk very tightly (more on this topic in another post)
