“Racorn-k”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
（→‎Introduction）
（→‎Experiments on several datasets）
 
（相同用户的63个中间修订版本未显示）
第1行： 第1行：
 
=Project name=
 
=Project name=
−
RACORN-K: RISK-AVERSION PATTERN MATCHING-BASED PORTFOLIO SELECTION
+
RACORN-K: Risk-aversion Pattern Matching-based Portfolio Selection
  
 
=Project members=
 
=Project members=
第24行： 第24行：
  
  
−
==Speaker feature learning==
+
==Corn-k==
 +
Corn-k algorithm is proposed by Li et al.[1].
 +
At the t-th trading period, the CORN-K algorithm first selects
 +
all the historical periods whose market status is similar to that
 +
of the present market, where the similarity is measured by
 +
the Pearson correlation coefficient. This patten matching process
 +
produces a set of similar periods, which we denote by C. Then do a optimization following the idea
 +
of BCRP[2] on C. Finally, the outputs of the top-k experts that have achieved the highest accumulated return are
 +
weighted to derive the ensemble-based portfolio.
  
−
The discovery of the short-time property of speaker traits is the key step towards speech signal factorization, as
+
==Racorn-k==
−
the speaker trait is one of the two main factors: the other is linguistic content that we have known for a long time
+
−
being short-time patterns.
+
  
−
The key idea of speaker feature learning is simply based on the idea of discriminating training speakers based on
+
The portfolio optimization is crucial for the success of
−
short-time frames by deep neural networks (DNN), date back to 2014 by Ehsan et al.[2]. As shown below, the output of the DNN
+
CORN-K. A potential problem of the existing form, however,
−
involves the training speakers, and the frame-level speaker features are read from the last hidden layer. The
+
is that the optimization is purely profit-driven. A natural idea
−
basic assumption here is: if the output of the last hidden layer can be used as the input feature of the
+
to consider the risk is to penalize risky portfolios when searching
−
last hidden layer (a software regression classifier), these features should be speaker discriminative.  
+
for the optimal portfolio. We use the standard deviation of log return
 +
on C to represent the risk and subtract this term as penalty.
  
−
[[文件:Dnn-spk.png|500px]]
+
==Racorn(c)-k==
  
−
However, the vanilla structure of Ehsan et al. performs rather poor compared to the i-vector counterpart. One reason is
+
The combination method used in CORN-K does not consider the time-variant property of the risk.
−
that the simple back-end scoring is based on average to derive the utterance-based representations (called d-vectors) , but
+
In fact, the risk of the portfolio derived from each expert tends to change quickly in an
−
another reason is the vanilla DNN structure that does not consider much of the context and pattern learning. We therefore
+
volatile market and therefore the weights of individual experts should be adjusted timely.
−
proposed a CT-DNN model that can learn stronger speaker features. The structure is shown below[1]:
+
To achieve the quick adjustment, we use the instant return s<sub>t</sub>(w,ρ,λ) to weight the
 +
experts with different λ, rather than the accumulated return.　Since s<sub>t</sub> is not available
 +
when estimating the optimal portfolio , we approximate it by the geometric average of the returns achieved in
 +
C.
  
−
[[文件:Ctdnn-spk.png|500px]]
+
=Experiments on several datasets=
  
 +
We evaluate our proposed algorithm on several dataset: DJIA, MSCI, SP500(N), HSI, SP500(O).
 +
DJIA, MSCI and SP500(O) are open dataset[3] that are used in previous work. you can find it
 +
[http://olps.stevenhoi.org/ here]. To observe the performance on more recent market, we
 +
collected another two dataset: SP500(N) and HSI
 +
([http://cslt.riit.tsinghua.edu.cn/mediawiki/images/9/9d/Sp500-n_hsi.rar download]).
 +
Form the following table, we can see that our algorithm can improve Sharpe ratio and reduce
 +
maximum drawdown consistently. In most case, it can also achieve better accumulated return.
  
−
Recently, we found that an 'all-info' training is effective for learning features. Looking back to DNN and CT-DNN, although the features
+
{| class="wikitable" style="margin: auto; text-align:center; width: 100%;"
−
read from last hidden layer are discriminative, but not 'all discriminative', because some discriminant info can be also impelemented
+
|+Performance summarization
−
in the last affine layer. A better strategy is let the feature generation net (feature net) learns all the things of discrimination.
+
|-
−
To achieve this, we discarded the parametric classifier (the last affine layer) and use the simple cosine distance to conduct the
+
|
−
classification. An iterative training scheme can be used to implement this idea, that is, after each epoch, averaging the speaker
+
|Dataset
−
features to derive speaker vectors, and then use the speaker vectors to replace the last hidden layer. The training will be then
+
|DJIA
−
taken as usual. The new structure is as follows[4]:
+
|MSCI
 +
|SP500(N)
 +
|HSI
 +
|SP500(O)
 +
|-
 +
|Criteria   
 +
|
 +
|RET  SR  MDD
 +
|RET  SR  MDD
 +
|RET  SR  MDD
 +
|RET  SR  MDD
 +
|RET  SR  MDD
 +
|-
 +
|rowspan="3"|Main Results
 +
|RACORN(C)-K 
 +
|0.93    0.01    0.32
 +
|78.38  3.73    0.21
 +
|12.55  0.77    0.53
 +
|202.04  1.60    0.28
 +
|7.13    1.28    0.34
 +
|-
 +
|RACORN-K 
 +
|0.83  -0.19 0.37
 +
|79.52  3.67 0.21
 +
|13.03  0.72 0.57
 +
|264.02 1.60 0.29
 +
|9.27  1.33 0.32
 +
|-
 +
|CORN-K   
 +
|0.80  -0.24 0.38
 +
|77.54  3.63 0.21
 +
|12.50  0.70 0.60
 +
|254.27 1.56 0.30
 +
|8.72  1.26 0.35
 +
|-
 +
|rowspan="2"|Naive Methods
 +
|UBAH   
 +
|0.76 -0.43 0.39
 +
|0.90  0.02 0.65
 +
|1.52  0.24 0.50
 +
|3.54  0.53 0.58
 +
|1.33  0.36 0.46
 +
|-
 +
|UCRP   
 +
|0.81 -0.28 0.38
 +
|0.92  0.05 0.64
 +
|1.78  0.28 0.68
 +
|4.25  0.58 0.55
 +
|1.64  0.55 0.31
 +
|-
 +
|rowspan="3"|Follow the Winner
 +
|UP     
 +
|0.81 -0.29 0.38
 +
|0.92  0.04 0.64
 +
|1.79  0.29 0.68
 +
|4.26  0.59 0.55
 +
|1.66  0.56 0.31
 +
|-
 +
|EG     
 +
|0.81 -0.29 0.38
 +
|0.92  0.04 0.64
 +
|1.75  0.28 0.67
 +
|4.22  0.58 0.55
 +
|1.62  0.54 0.32
 +
|-
 +
|ONS   
 +
|1.53 0.80 0.32
 +
|0.85 0.02 0.68
 +
|0.78 0.27 0.96
 +
|4.42 0.52 0.68
 +
|3.32 1.11 0.25
 +
|-
 +
|rowspan="6"|Follow the Loser
 +
|ANTICOR   
 +
|1.62  0.85 0.34
 +
|2.75  0.96 0.51
 +
|1.16  0.24 0.93
 +
|9.10  0.74 0.56
 +
|5.58  1.08 0.38
 +
|-
 +
|ANTICOR2 
 +
|2.28  1.24 0.35
 +
|3.20  1.02 0.48
 +
|0.71  0.22 0.97
 +
|12.27 0.77 0.55
 +
|5.86  1.01 0.49
 +
|-
 +
|PAMR2   
 +
|0.70 -0.15 0.76
 +
|16.73 2.07 0.54
 +
|0.01 -0.28 1.00
 +
|1.19 0.20 0.86
 +
|4.97 0.90 0.51
 +
|-
 +
|CWMR Stdev 
 +
|0.69 -0.17 0.76
 +
|17.14 2.07 0.54
 +
|0.02 -0.26 0.99
 +
|1.28 0.22 0.85
 +
|5.92 0.96 0.51
 +
|-
 +
|OLMAR1   
 +
|2.53  1.16 0.37
 +
|14.82 1.85 0.48
 +
|0.03 -0.11 1.00
 +
|4.19 0.46 0.77
 +
|15.89 1.28 0.41
 +
|-
 +
|OLMAR2   
 +
|1.16  0.40 0.58
 +
|22.34 2.08 0.42
 +
|0.03 -0.11 1.00
 +
|3.65  0.43 0.84
 +
|9.54  1.08 0.49
 +
|-
 +
|rowspan="2"|Pattern Matching
 +
|BK     
 +
|0.69 -0.68 0.43
 +
|2.62 1.06 0.51
 +
|1.97  0.31 0.59
 +
|13.90 0.88 0.45
 +
|2.21 0.64 0.49
 +
|-
 +
|BNN   
 +
|0.88 -0.15 0.31
 +
|13.40 2.33 0.33
 +
|6.81  0.67 0.41
 +
|104.97 1.40 0.33
 +
|3.05 0.76 0.45
 +
|}
  
 +
=View the impact of risk-aversion=
  
−
[[文件:fullinfo-spk.png|500px]]
+
From RET curves we can see that RACORN(C)-K behaves more
 +
‘smooth’ than CORN-K. Due to this smoothness, the risk
 +
of the strategy is reduced, and extremely poor trading can be
 +
largely avoided.
  
−
=Speech factorization=
+
[[文件:Racornk-fig1.png|600px]]
  
  
−
The short-time property is a very nice thing, which tells us it is possible to factorize speech signals. By factorization, we can achieve significant benefits:
+
==View the impact of risk-aversion on volatile markets==
  
−
A. Individual tasks can be largely improved, as unrelated factors have been removed.
+
On volatile markets, our proposed algorithms are more effective.
−
 
+
−
B. Factors that are disturbs now becomes valuables things, leading to conditional training and collaborative training [5].
+
−
 
+
−
C. Once the factors have been separated, single factors can be manipulated, and reassemble these factors can change the signal according to the need.
+
−
 
+
−
D. It is a new speech coding scheme that leverage knowledge learned from large data.
+
−
 
+
−
 
+
−
Traditional factorization methods are based on probabilistic models and maximum likelihood learning. For example, in JFA, a linear Gaussian is assumed
+
−
for speaker and channel, and then a ML estimation is applied to estimate the loading matrices of each factor, based on a long duration of speech. Almost
+
−
all these factorizations share these features: shallow, linear, Gaussian, long-term segments.
+
−
 
+
−
We are more interested in factorization on frame-level, and plays not much assumption on how the factors are mixed. A cascaded factorization approach has been
+
−
proposed[6]. The basic idea is to factorize significant factors first,and then conditioned on the factors that have been derived. The architecture is as
+
−
follows, where we factorized speech signals into three factors: linguistic content, speaker trait, emotion. When factorizing each factor, supervised
+
−
learning is used. Note by this architecture, databases with different target labels can be used in a complementary way, which is different from
+
−
previously joint training approach that needs full-labelled data.
+
−
 
+
−
 
+
−
[[文件:deepfact.png|500px]]
+
−
 
+
−
 
+
−
=Speech reconstruction=
+
−
 
+
−
To verify the factorization, we can reconstruct the speech signal from the factors. The reconstruction is simply based on a DNN,
+
−
as shown below.
+
−
Each factor passes a unique deep neural net, the output of the three DNNs are added together, and compared with the target,
+
−
which is the logarithm of the spectrum of the original signal. This means that the output of the DNNs of the three factors are
+
−
assumed to be convolved together to produce the original speech.
+
−
 
+
−
[[文件:fact-recover-dnn.png|500px]]
+
−
 
+
−
Note that the factors are learned from Fbanks, by which some speech information
+
−
has been lost, however the recovery is rather successfull.
+
−
 
+
−
 
+
−
==View the reconstruction==
+
−
 
+
−
[[文件:fact-recover.png|500px]]
+
−
 
+
−
 
+
−
More recovery examples can be found [[dsf-examples|here]].
+
−
 
+
−
 
+
−
==Listen to the reconstruction==
+
−
 
+
−
We can listen to the wave for each factor, by using the original phase.
+
−
 
+
−
 
+
−
Original speech: [http://wangd.cslt.org/research/cdf/demo/CHEAVD_1_1_E02_001_worried.wav]
+
−
 
+
−
Linguistic factor: [http://wangd.cslt.org/research/cdf/demo/phone.wav]
+
−
 
+
−
Speaker factor: [http://wangd.cslt.org/research/cdf/demo/speaker.wav]
+
−
 
+
−
Emotion factor: [http://wangd.cslt.org/research/cdf/demo/emotion.wav]
+
−
 
+
−
Liguistic+ Speaker + Emotion: [http://wangd.cslt.org/research/cdf/demo/recovery.wav]
+
−
 
+
−
 
+
−
=Research directions=
+
−
 
+
−
* Adversarial factor learning
+
−
* Phone-aware multiple d-vector back-end for speaker recognition
+
−
* TTS adaptation based on speaker factors
+
  
 +
[[文件:Racornk-fig2.png|600px]]
  
 
=Reference=
 
=Reference=
  
−
[1] Lantian Li, Yixiang Chen, Ying Shi, Zhiyuan Tang, and Dong Wang, “Deep speaker feature learning for text-independent speaker verification,”, Interspeech 2017.  
+
[1] Bin Li, Steven CH Hoi, and Vivekanand Gopalkrishnan, “Corn: Correlation-driven nonparametric learning approach for portfolio selection,”, ACM Transactions on Intelligent Systems and Technology (TIST), vol. 2, no. 3,pp. 21, 2011.  
−
 
+
−
[2] Ehsan Variani, Xin Lei, Erik McDermott, Ignacio Lopez Moreno, and Javier Gonzalez-Dominguez, “Deep neural networks for small footprint text-dependent speaker
+
−
verification,”, ICASSP 2014.
+
−
 
+
−
[3] Lantian Li, Dong Wang, Yixiang Chen, Ying Shing, Zhiyuan Tang, http://wangd.cslt.org/public/pdf/spkfact.pdf
+
  
−
[4] Lantian Li, Zhiyuan Tang, Dong Wang, FULL-INFO TRAINING FOR DEEP SPEAKER FEATURE LEARNING, http://wangd.cslt.org/public/pdf/mlspk.pdf
+
[2] Thomas M Cover and David H Gluss, “Empirical bayes stock market portfolios,” Advances in applied mathematics, vol. 7, no. 2, pp. 170–181, 1986.
  
−
[5] Zhiyuan Thang, Lantian Li, Dong Wang, Ravi Vipperla "Collaborative Joint Training with Multi-task Recurrent Model for Speech and Speaker Recognition", IEEE Trans. on Audio, Speech and Language Processing, vol. 25, no.3, March 2017.
+
[3] B Li, D Sahoo, and SCH Hoi, “Olps: A toolbox for online portfolio selection,” Journal of Machine Learning Research (JMLR), 2015.
  
−
[6] Dong Wang,Lantian Li,Ying Shi,Yixiang Chen,Zhiyuan Tang., "Deep Factorization for Speech Signal", https://arxiv.org/abs/1706.01777
+
=Contact Me=
 +
Email: yang-wang16@mails.tsinghua.edu.cn

2018年3月17日 (六) 11:37的最后版本

Project name

RACORN-K: Risk-aversion Pattern Matching-based Portfolio Selection

Project members

Yang Wang, Dong Wang, Yaodong Wang, You Zhang

Introduction

Portfolio selection is the central task for assets management, but it turns out to be very challenging. Methods based on pattern matching, particularly the CORN-K algorithm, have achieved promising performance on several stock markets. A key shortage of the existing pattern matching methods, however, is that the risk is largely ignored when optimizing portfolios, which may lead to unreliable profits, particularly in volatile markets. To make up this shortcoming， We propose a risk-aversion CORN-K algorithm, RACORN-K, that penalizes risk when searching for optimal portfolios. Experiment results demonstrate that the new algorithm can deliver notable and reliable improvements in terms of return, Sharp ratio and maximum drawdown, especially on volatile markets.


Corn-k

Corn-k algorithm is proposed by Li et al.[1]. At the t-th trading period, the CORN-K algorithm first selects all the historical periods whose market status is similar to that of the present market, where the similarity is measured by the Pearson correlation coefficient. This patten matching process produces a set of similar periods, which we denote by C. Then do a optimization following the idea of BCRP[2] on C. Finally, the outputs of the top-k experts that have achieved the highest accumulated return are weighted to derive the ensemble-based portfolio.

Racorn-k

The portfolio optimization is crucial for the success of CORN-K. A potential problem of the existing form, however, is that the optimization is purely profit-driven. A natural idea to consider the risk is to penalize risky portfolios when searching for the optimal portfolio. We use the standard deviation of log return on C to represent the risk and subtract this term as penalty.

Racorn(c)-k

The combination method used in CORN-K does not consider the time-variant property of the risk. In fact, the risk of the portfolio derived from each expert tends to change quickly in an volatile market and therefore the weights of individual experts should be adjusted timely. To achieve the quick adjustment, we use the instant return st(w,ρ,λ) to weight the experts with different λ, rather than the accumulated return.　Since st is not available when estimating the optimal portfolio , we approximate it by the geometric average of the returns achieved in C.

Experiments on several datasets

We evaluate our proposed algorithm on several dataset: DJIA, MSCI, SP500(N), HSI, SP500(O). DJIA, MSCI and SP500(O) are open dataset[3] that are used in previous work. you can find it here. To observe the performance on more recent market, we collected another two dataset: SP500(N) and HSI (download). Form the following table, we can see that our algorithm can improve Sharpe ratio and reduce maximum drawdown consistently. In most case, it can also achieve better accumulated return.

Performance summarization
Dataset DJIA MSCI SP500(N) HSI SP500(O)
Criteria RET SR MDD RET SR MDD RET SR MDD RET SR MDD RET SR MDD
Main Results RACORN(C)-K 0.93 0.01 0.32 78.38 3.73 0.21 12.55 0.77 0.53 202.04 1.60 0.28 7.13 1.28 0.34
RACORN-K 0.83 -0.19 0.37 79.52 3.67 0.21 13.03 0.72 0.57 264.02 1.60 0.29 9.27 1.33 0.32
CORN-K 0.80 -0.24 0.38 77.54 3.63 0.21 12.50 0.70 0.60 254.27 1.56 0.30 8.72 1.26 0.35
Naive Methods UBAH 0.76 -0.43 0.39 0.90 0.02 0.65 1.52 0.24 0.50 3.54 0.53 0.58 1.33 0.36 0.46
UCRP 0.81 -0.28 0.38 0.92 0.05 0.64 1.78 0.28 0.68 4.25 0.58 0.55 1.64 0.55 0.31
Follow the Winner UP 0.81 -0.29 0.38 0.92 0.04 0.64 1.79 0.29 0.68 4.26 0.59 0.55 1.66 0.56 0.31
EG 0.81 -0.29 0.38 0.92 0.04 0.64 1.75 0.28 0.67 4.22 0.58 0.55 1.62 0.54 0.32
ONS 1.53 0.80 0.32 0.85 0.02 0.68 0.78 0.27 0.96 4.42 0.52 0.68 3.32 1.11 0.25
Follow the Loser ANTICOR 1.62 0.85 0.34 2.75 0.96 0.51 1.16 0.24 0.93 9.10 0.74 0.56 5.58 1.08 0.38
ANTICOR2 2.28 1.24 0.35 3.20 1.02 0.48 0.71 0.22 0.97 12.27 0.77 0.55 5.86 1.01 0.49
PAMR2 0.70 -0.15 0.76 16.73 2.07 0.54 0.01 -0.28 1.00 1.19 0.20 0.86 4.97 0.90 0.51
CWMR Stdev 0.69 -0.17 0.76 17.14 2.07 0.54 0.02 -0.26 0.99 1.28 0.22 0.85 5.92 0.96 0.51
OLMAR1 2.53 1.16 0.37 14.82 1.85 0.48 0.03 -0.11 1.00 4.19 0.46 0.77 15.89 1.28 0.41
OLMAR2 1.16 0.40 0.58 22.34 2.08 0.42 0.03 -0.11 1.00 3.65 0.43 0.84 9.54 1.08 0.49
Pattern Matching BK 0.69 -0.68 0.43 2.62 1.06 0.51 1.97 0.31 0.59 13.90 0.88 0.45 2.21 0.64 0.49
BNN 0.88 -0.15 0.31 13.40 2.33 0.33 6.81 0.67 0.41 104.97 1.40 0.33 3.05 0.76 0.45

View the impact of risk-aversion

From RET curves we can see that RACORN(C)-K behaves more ‘smooth’ than CORN-K. Due to this smoothness, the risk of the strategy is reduced, and extremely poor trading can be largely avoided.

Racornk-fig1.png


View the impact of risk-aversion on volatile markets

On volatile markets, our proposed algorithms are more effective.

Racornk-fig2.png

Reference

[1] Bin Li, Steven CH Hoi, and Vivekanand Gopalkrishnan, “Corn: Correlation-driven nonparametric learning approach for portfolio selection,”, ACM Transactions on Intelligent Systems and Technology (TIST), vol. 2, no. 3,pp. 21, 2011.

[2] Thomas M Cover and David H Gluss, “Empirical bayes stock market portfolios,” Advances in applied mathematics, vol. 7, no. 2, pp. 170–181, 1986.

[3] B Li, D Sahoo, and SCH Hoi, “Olps: A toolbox for online portfolio selection,” Journal of Machine Learning Research (JMLR), 2015.

Contact Me

Email: yang-wang16@mails.tsinghua.edu.cn

取自“http://index.cslt.org/mediawiki/index.php?title=Racorn-k&oldid=29903”