0 penilaian0% menganggap dokumen ini bermanfaat (0 suara)
183 tayangan11 halaman
In this paper, we create an aggregate filter based upon RavenPack’s news data to alleviate this problem. We compare it to using VIX as a filter. We find that our aggregated RavenPack based filter has risk adjusted returns of 0.84 compared with 0.41 for a VIX filter and 0.47 for a long only strategy when trading G10 FX carry since 2002. At the same time, the news filter also cuts FX carry drawdowns in half. Our RavenPack based filter also reduces drawdowns for long only S&P500 futures strategies from 57.1% to 26%, whilst increasing risk adjusted returns from 0.31 to 0.44.
In this paper, we create an aggregate filter based upon RavenPack’s news data to alleviate this problem. We compare it to using VIX as a filter. We find that our aggregated RavenPack based filter has risk adjusted returns of 0.84 compared with 0.41 for a VIX filter and 0.47 for a long only strategy when trading G10 FX carry since 2002. At the same time, the news filter also cuts FX carry drawdowns in half. Our RavenPack based filter also reduces drawdowns for long only S&P500 futures strategies from 57.1% to 26%, whilst increasing risk adjusted returns from 0.31 to 0.44.
In this paper, we create an aggregate filter based upon RavenPack’s news data to alleviate this problem. We compare it to using VIX as a filter. We find that our aggregated RavenPack based filter has risk adjusted returns of 0.84 compared with 0.41 for a VIX filter and 0.47 for a long only strategy when trading G10 FX carry since 2002. At the same time, the news filter also cuts FX carry drawdowns in half. Our RavenPack based filter also reduces drawdowns for long only S&P500 futures strategies from 57.1% to 26%, whilst increasing risk adjusted returns from 0.31 to 0.44.
1 Thalesians Ltd. Non-independent investment research (see disclaimers)
Carry the news trade Filtering FX carry using RavenPack news analytics 2 Sep 2014 Over the decade that I have created systematic FX models, carry has always been cited as one of the most popular strategies (and in the years preceding that too). Creating a generic carry basket is relatively simple and has offered reasonable historical returns for collecting a risk premium. However, FX carry suffers from drawdowns during times of risk aversion, when investors seek to cut their exposure in risky assets. In this paper, we create an aggregate filter based upon RavenPacks news data to alleviate this problem. We compare it to using VIX as a filter. We find that our aggregated RavenPack based filter has risk adjusted returns of 0.84 compared with 0.41 for a VIX filter and 0.47 for a long only strategy when trading G10 FX carry since 2002. At the same time, the news filter also cuts FX carry drawdowns in half. Our RavenPack based filter also reduces drawdowns for long only S&P500 futures strategies from 57.1% to 26%, whilst increasing risk adjusted returns from 0.31 to 0.44. This paper has been kindly sponsored by RavenPack, a pioneer in financial news and sentiment analytics. Please contact saeed@thalesians.com if you are interested in learning about our quant consulting services and more about our research at the Thalesians. Also see http://www.thalesians.com and follow us on Twitter @thalesians. Introduction Whilst historically having long exposure in risky assets such as S&P500 and FX carry has been profitable, it has been accompanied by drawdowns during periods of risk aversion. In this paper, we investigate how RavenPack news data can be used to create a filter for risky assets. In Figure 1, we show how it can be used to improve the risk adjusted returns for G10 FX carry and reduce drawdowns by more than half, when compared to long only FX carry exposure or when using a VIX filter. Later, we shall also demonstrate how it can be used to trade S&P500. Figure 1: Using Aggregated RavenPack news signal to filter G10 FX carry basket
Source: Thalesians, RavenPack, Bloomberg
As we remarked earlier, FX carry has been a popular strategy in FX for many years. Indeed, using an extended sample going back to the 1970s, using developed market currencies, we can see that FX carry has broadly been profitable (see Figure 2). Essentially, FX carry involves buying higher yielding currencies and funding these purchases through selling low 70 120 170 220 2002 2004 2006 2008 2010 2012 2014 No filter Ret=5.3% Vol=11.3% IR=0.47 Dr=-35.5% VIX Ret=3.5% Vol=8.4% IR=0.41 Dr=-27.8% Agg RavenPack Ret=5.3% Vol=6.3% IR=0.84 Dr=-15% Saeed Amen Quantitative Strategy +44 20 3290 9624 saeed@thalesians.com @thalesians http://www.thalesians.com
THE THALESIANS Cross Asset / Quant Strategy
2 Thalesians Ltd. Non-independent investment research (see disclaimers)
yielding currencies. In our case, we have simply bought the four highest yielding currencies in G10 and sold the four lowest yielding currencies, reweighting our basket on a monthly basis. Figure 2: G10 FX carry basket (RHS) has broadly been profitable since the 1970s vs. S&P500 (LHS)
Source: Thalesians, Bloomberg
A carry investor collects the carry differential between the high yielding and low yielding currencies in his or her carry basket. The trade is profitable provided there are not large depreciations of the high yielding currencies vs. the lower yielding currencies in the basket. Unfortunately, carry trades can be subject to large drawdowns during periods of risk aversion when investors priorities shift from seeking return from higher yielding assets to safe guarding capital. Indeed, we note the significant drawdowns in carry have often coincided with those in equities, notably in 2008 (see Figure 2). In Figure 3, we present the total returns of being long AUD/JPY, which is a classic carry trade. Generally Australian yields have been higher than Japanese yields. Hence, a carry trader would generally be long AUD/JPY throughout history. We note that total returns (which include carry) have been much higher over time than spot returns. The difference between the two is the carry, which actually constitutes the bulk of returns. Hence, we see that over a long period of time a large amount of carry can accrue. Thus, if we wish to profit from a carry trade, we need to be holding it for a long period of time.
3 Thalesians Ltd. Non-independent investment research (see disclaimers)
Figure 3: Comparing total & spot returns for long AUD/JPY strategies
Source: Thalesians, Bloomberg
The main challenge facing FX carry traders is finding methods of reducing drawdowns associated with the strategy. Typically these involve filters which give an indication of risk sentiment. When the indicator signals risk aversion the carry basket is usually exited. In this paper, we shall investigate a filter which is based upon RavenPack news data. We shall compare it to using VIX as a risk filter which tends to be one of the most prevalent ways to filter exposure to risky assets such as FX carry and S&P500. Outlining steps for creating a news based trading rule Before we discuss any sort of trading rule, we need to understand the general ideas behind news analysis. The process behind trading news data is more involved than the process of creating a carry basket, which is somewhat simpler. We note that there are many steps involved in analysing news data for trading purposes, which we list below. 1. Analysing text from news articles/headlines (RavenPack does this step!) 2. Aggregate RavenPack data by time frequency 3. Create a sentiment index for specific areas of interest 4. Apply a trading rule to the sentiment index Next, we shall endeavour to discuss each step separately in some detail. 1 / Analysing text from news articles/headlines The most complex step comes from initially analysing the text in news. RavenPack essentially does this step for us. RavenPack products are available for three different data sources, which we describe below. Dow Jones Edition Dow Jones Newswires, regional editions of the Wall Street Journal and Barron's Web Edition Business publishers, national and local news, blog sites, government and regulatory updates 19,000 different sources PR Edition 22 newswires and press release distribution networks More than 100,000 press releases and regulatory disclosures processed every day In our analysis, we shall be using the Full Edition which is an aggregation of Dow Jones, Web and PR Editions. Our focus is on RavenPacks new Macro 4.0 product 1 . Using proprietary
1 There is also a product which covers equities. 0 50 100 150 200 250 300 1995 2000 2005 2010 AUDJPY Spot Ret=2.53% Vol=17.02% IR=0.15 Dr=-48.26% AUDJPY ToT Ret=6.28% Vol=16.97% IR=0.37 Dr=-46.69%
THE THALESIANS Cross Asset / Quant Strategy
4 Thalesians Ltd. Non-independent investment research (see disclaimers)
technology, RavenPack classify news events from these various sources. For each news event analysed, a record is generated which includes 34 fields. Below, we give a small selection of the fields recorded for each news event. Timestamp of publication In UTC time with a millisecond timestamp Focus of the publication Includes details on the country and the general subject of the news We shall use these fields later to filter news for example for US news related to the economy Positive/negative nature of news Scaled from 0 to 100, where >50 is positive, <50 is negative (and 50 is neutral) We shall use this later for identifying the bullishness/bearishness of an article for trading purposes (Event sentiment score ESS) Measures of the relative novelty of news Newer news as opposed to repeated headlines scores higher (Event Novelty score ENS) Prevalence of news Identify the number of positive or negative events for a certain entity (Aggregate Event Sentiment AES) and also the general news volume on an entity (Aggregate Event Volume AEV) Source of the news and the RavenPack product edition The RavenPack News Analytics guide has a much more exhaustive description of the fields and I would strongly recommend reading that before attempting to use the data 2 . Given the very detailed nature of every news record, the user has a lot of flexibility in how they interpret every news record. Traders are free to use as few or as many of the fields from each news record in their analysis. As a result, it is likely that the final trading algorithms will differ somewhat between the various traders who use this news analytics data. This reduces the likelihood of crowding around certain signals. 2 / Aggregate RavenPack data by time frequency When the data is received from RavenPack, the trader needs to aggregate it according to time frequency. For example, a high frequency trader might wish to continually monitor the feed, reading each record separately, as it arrives in the feed and creating a trading rule which can be triggered on each read. The news data can be accessed in realtime via an API for intraday trading or via daily and monthly CSV files for longer term trading. In our case, given we are using the news data to trade markets on a daily basis at the close, we first aggregate it on an hourly basis. From that we create a daily aggregation of the news data, which cuts off at 4pm NY every day giving us a raw feed. This is before the closing snapshots of our traded assets (NY close). Clearly, we must be careful that our trading signal is not generated after our trading snapshot. Having intraday RavenPack data makes this somewhat easier. 3 / Create indices for specific areas of interest Given we now have the news data sorted into daily chunks, we need to decide the focus of our news data search. Our intension is to create a filter for FX carry (and also for risky assets more broadly). Hence, as an approximation we shall focus on US based data. As the largest economy in the world, the US obviously has a large impact on global sentiment. In practice,
2 This will save you a lot of time!
THE THALESIANS Cross Asset / Quant Strategy
5 Thalesians Ltd. Non-independent investment research (see disclaimers)
we might wish to create broader based indices that cover more countries (and perhaps GDP weight them). As with any trading strategy, simply throwing more data at the problem is not necessarily the answer. It is more a case of using the right dataset and using a more directed search approach using some underlying rationale. Indeed, it is this approach which we seek to adopt throughout this paper. Volumebasedmeasure First, we create a volume based index based on the number of news articles, which we describe below. US newscount Total number of articles about US which we use as a proxy for market volatility. We are interested in market volatility, because higher levels of market volatility are associated with more market risk aversion. These are precisely the times, when high beta trades (such as FX carry and also long S&P500) fall out of fashion, which we mentioned in the introduction. In Figure 4, we plot our raw aggregated US newscount against VIX (a measure of equity volatility). We note the massive spike of VIX during the Lehman crisis in 2008. Furthermore, we see similar spikes during mid-2010, during the Greek debt crisis. At least from a stylistic point of view, it appears that there is a strong connection between market volatility and risk aversion. If we shift our attention to the raw US newscount measure, we see that it increases over time. This should not be surprising. From 2007 onwards, we notice an acceleration in the rise (recall that the RavenPack Web Edition starts in 2007). Furthermore, the signal appears quite noisy. Hence, we need to do some work on the raw US newscount in order to create a trading signal, which includes some de-noising of the initial data. We also need to consider some element of seasonality in the initial data. We create a scaled measure which does some adjustment for the progressive increase in news volume over time. We also create a scaled measure for VIX as well. These scaled measures enable us to identify localised spikes in either VIX or US newscounts. These can be used as inputs into our trading rule. We plot both these measures in Figure 5. Figure 4: Raw US newscount vs. VIX Figure 5: Scaled US news vs. VIX score
6 Thalesians Ltd. Non-independent investment research (see disclaimers)
Pressureindices As well as our US newscount based score, which measures news volume as discussed above, we also construct our own news pressure indices based upon the news data. These are related to the five topics identified by RavenPack. These topics are used to classify some of the news events. Our pressure indices give a relative positive/negative score, based upon an aggregation of RavenPacks ESS 3 (event sentiment score). US business US economy US environment US politics US society In Figure 6, we plot the US business pressure index, whilst in Figure 7, we plot the US economy pressure index. We show the raw aggregated indices against the pressure indices. We see that the initial data is considerably noisier than the pressure indices. It is important to de-noise the initial data, otherwise, it will generate too many trading signals for our purposes (as we did with the US newscount raw data). Furthermore, the relative magnitudes of the various raw indices we have constructed are quite different. We have adjusted for this when constructing the pressure indices. We have adopted an identical process for the construction of the various pressure indices, rather than fitting each index separately. Figure 6: US business pressure index Figure 7: US economy pressure index
Source: Thalesians, RavenPack
Source: Thalesians, RavenPack
We also create a sub-index of the economy topic, using the US employment group. In Figure 8, we plot the US employment pressure index against the changes in the change in US nonfarm payrolls. In other words, we are looking at the difference in the US nonfarm payrolls number this month versus last month. We note that there is some relationship between the two time series. We are using the revised payrolls series in the plot, which in practice, is only available with a further one month lag. Hence, any trading signal would also have to be lagged, if were to use the revised payroll data. This of course contrasts to
3 The ESS (Event Sentiment Score) is available for those news events which have classified. News events which have not been classified are not given ESSs. 0 20 40 60 80 -3 -2 -1 0 1 2 3 Jan 2008 Jul 2008 Jan 2009 Jul 2009 Jan 2010 US business pressure US business raw Pressure Raw 0 20 40 60 80 -3 -2 -1 0 1 2 3 Jan 2008 Jul 2008 Jan 2009 Jul 2009 Jan 2010 US economy pressure US economy raw Pressure Raw
THE THALESIANS Cross Asset / Quant Strategy
7 Thalesians Ltd. Non-independent investment research (see disclaimers)
our pressure index, which we could even calculate in real-time if we were accessing the news data via the API. Figure 8: US employment pressure index against changes in changes in US NFP
Source: Thalesians, RavenPack, Bloomberg
Of course, we could have constructed a large number of different indices using RavenPack data which would also likely be useful (for example indices based upon news related to specific assets, such as specific currencies). We hope to explore the data more in the future, to find other approaches to extract different trading signals from the data. 4 / Apply a trading rule to the volume and pressure indices We have now developed a multitude of US newscount and news pressure indices, based upon RavenPack news data. Later, we shall use these to filter long G10 FX carry and long S&P500 futures exposure. For US newscount and VIX, we adopt a simple cut off signal based upon their associated scores, which we describe below: If the score is high (above 0.75), we exit FX carry and stay flat If the score is low (below 0.75), we remain long FX carry Our US pressure indices however, give us relative positive and negative measures, which we utilise in the trading signals, as below. If the pressure index is positive, buy FX carry If the pressure index is negative, go flat FX carry Whilst the indices themselves are relatively complex to construct, given the number of steps we have taken, we note that the final trading rules are actually fairly straightforward. Backtesting our trading rules In this section, we backtest our trading rules as filters for a G10 FX carry basket. Our data set is from 2002 and we include both transaction costs and carry in our analysis. In Figures 9 and 10, we present the backtested return statistics for trading FX carry with these various trading rules. Our first observation is that without a filter the drawdowns are the worst, closely followed by the VIX based filter. We note that those information ratios based upon US newscount, US economy and employment pressure indices are significantly higher than those for no filter. This seems to intuitively make sense, namely that better US macro-economic news (and lower vol) are positive for carry. The US politics and society pressure indices also perform strongly for -400 -300 -200 -100 0 100 200 300 400 -4 -3 -2 -1 0 1 2 3 4 Jan 2008 Jul 2008 Jan 2009 Jul 2009 Jan 2010 US employment pressure US payrolls (difference from last) Pressure K jobs
THE THALESIANS Cross Asset / Quant Strategy
8 Thalesians Ltd. Non-independent investment research (see disclaimers)
filtering carry. Elsewhere, the US business and environment filters have lower risk adjusted returns, but they still have better drawdowns than using no filter. Further work could include stripping down the US economy topic into more of its constituent parts, as we have done for employment. This would enable us to better understand the various groups of data which are included in the US economy topic (and remove ones which are less relevant). Figure 9: FX carry risk adjusted returns Figure 10: FX carry drawdowns
Source: Thalesians, RavenPack, Bloomberg
Source: Thalesians, RavenPack, Bloomberg
As a next step we create an aggregate RavenPack signal, which is an equally weighted portfolio of US newscount, US economy and US employment filtered baskets. Hence the aggregate RavenPack signal captures both market volatility and the US macro-economic situation. We plot the cumulative returns of the aggregated RavenPack signal, alongside a VIX filtered and unfiltered FX carry basket in Figure 11. Year-on-year returns are presented in Figure 12. We note that our aggregated signal using RavenPack data has an information ratio of 0.84 and drawdowns of 15%. This contrasts to an information ratio of 0.47 and a drawdown of 35.5% for the unfiltered version. Figure 11: FX carry cumulative Figure 12: FX carry YoY
Source: Thalesians, RavenPack, Bloomberg
Source: Thalesians, RavenPack, Bloomberg
We repeat the exercise for filtering S&P500 futures, using the aggregated RavenPack filter, comparing it to a VIX filter and also using no filter. We plot cumulative returns of these strategies in Figures 13 and 14. 0.0 0.2 0.4 0.6 0.8 1.0 U S
9 Thalesians Ltd. Non-independent investment research (see disclaimers)
We note that our generic VIX filter performs relatively poorly. The unfiltered strategy (in other words long only S&P500 futures) performs better and has positive risk adjusted return at 0.31. However, the drawdowns are significant at 57.1%. By contrast the aggregated RavenPack news filtered strategy, has a higher risk adjusted return at 0.44 and the drawdowns are nearly half at 26.1%. Figure 13: S&P500 cumulative Figure 14: S&P500 YoY
Source: Thalesians, RavenPack, Bloomberg
Source: Thalesians, RavenPack, Bloomberg
Conclusion We have given an introduction to the idea of G10 FX carry. We have noted that whilst it is profitable, it can be subject to drawdowns during periods of market risk aversion. We described how to construct trading signals from RavenPack news data, in particular for the filtering of FX carry strategies. Our aggregated RavenPack news filter considerably improved risk adjusted returns for a G10 FX carry basket, when compared to an unfiltered basket (or a generic VIX filter). Indeed, our aggregated RavenPack based filter has risk adjusted returns of 0.84 compared with 0.41 for a VIX filter or 0.47 for an unfiltered strategy when trading G10 FX carry since 2002. Furthermore, the aggregated RavenPack filter cut drawdowns for FX carry in half, compared to either an unfiltered or VIX filtered strategy. Later, we also showed how the aggregated RavenPack based filter reduced drawdowns for long-only S&P500 futures strategies from 57.1% to 26.1%, whilst increasing risk adjusted returns from 0.31 to 0.44. This suggests that the aggregated RavenPack filter we have created likely has applications to more broadly filter long high beta strategies elsewhere. For more information about RavenPack News Analytics data, please feel free to contact saeed@thalesians.com (and see section below) or if you would like to use our consulting services to interpret RavenPack News Analytics data to create systematic trading models. About RavenPack Data RavenPack News Analytics (RPNA) provides real-time structured sentiment, relevance and novelty data for entities and events detected in unstructured text published by reputable sources. Publishers include Dow Jones Newswires, Barrons, the Wall Street Journal and over 22,000 other traditional and social media sites. Up to 14 years of Dow Jones newswires archive and 7 years of historical data from web publications and blogs are available for backtesting. 40 100 160 220 2002 2004 2006 2008 2010 2012 2014 No filter Ret=6.4% Vol=20.8% IR=0.31 Dr=-57.1% VIX Ret=1% Vol=14.7% IR=0.07 Dr=-55.1% Agg RavenPack Ret=5.1% Vol=11.6% IR=0.44 Dr=- 26.1% -60% -40% -20% 0% 20% 40% 2 0 0 2 2 0 0 3 2 0 0 4 2 0 0 5 2 0 0 6 2 0 0 7 2 0 0 8 2 0 0 9 2 0 1 0 2 0 1 1 2 0 1 2 2 0 1 3 2 0 1 4 No filter VIX Agg RavenPack
THE THALESIANS Cross Asset / Quant Strategy
10 Thalesians Ltd. Non-independent investment research (see disclaimers)
RavenPack detects news and produces analytics data on over 33,000 listed stocks from the world's equity markets, over 2,300 financially relevant organizations, 138,000 places, 150 currencies and 80 commodities.
THE THALESIANS Cross Asset / Quant Strategy
11 Thalesians Ltd. Non-independent investment research (see disclaimers)
Non-independent investment research disclaimer This investment research has not been prepared in accordance with legal requirements intended to promote the independence of investment research. It is also not subject to any prohibition on dealing ahead of the dissemination of investment research. Thalesians Ltd., its affiliates or staff, may perform services for, solicit business from, hold long or short positions in, or otherwise be interested in the investments (including derivatives), of any issuer mentioned herein. Material within this note is confidential and should not be copied, distributed, published or reproduced in whole or in part or disclosed by recipients to any other person. Any information or opinions in this material are not intended for distribution to, or use by, any person in any jurisdiction or country where such distribution or use would be unlawful. It is intended purely for the consumption of professional investors. Thalesians Ltd. does not warrant or represent that the document is accurate, complete, reliable, fit for any particular purpose or merchantable and does not accept liability for any act (or decision not to act) resulting from use of this document and related data. To the maximum extent permissible all warranties and other assurances by Thalesians Ltd. are hereby excluded and Thalesians Ltd. shall have no liability for the use, misuse, or distribution of this information. Past investment performance is no indication of future investment performance. Thalesians Ltd., PO Box 309, 56 Gloucester Road, London SW7 AUB, UK Tel +44 20 3290 9624, e-mail saeed@thalesians.com web http://www.thalesians.com Thalesians Ltd. is registered as a company in UK with company no. 06843387 Copyright 2014 Thalesians Ltd. All rights reserved