Estimating Models Based on Markov Jump Processes Given Fragmented Observation Series Markus Hahn Johann Radon Institute for Computational and Applied Mathematics (RICAM), Linz Austrian Academy of Sciences Joint work with S. Fr¨ uhwirth-Schnatter (JKU Linz) and J. Sass (TU Kaiserslautern) Linz, December 2, 2008 Work funded by FWF project P17947
37
Embed
Estimating Models Based on Markov Jump Processes Given ... · Estimating Models Based on Markov Jump Processes Given Fragmented Observation ... Markov Switching ... Markov Jump Processes
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Estimating Models Based on Markov JumpProcesses Given Fragmented Observation Series
Markus Hahn
Johann Radon Institute for Computational and Applied Mathematics (RICAM), LinzAustrian Academy of Sciences
Joint work with S. Fruhwirth-Schnatter (JKU Linz) and J. Sass (TU Kaiserslautern)
Linz, December 2, 2008
Work funded by FWF project P17947
Introduction
Problem
I Estimation from set of observed seriesI Independent series
I Data with breaks
I Each single series is (based) Markov processI Same generator for all series
I How to estimate common generator?
I How to cope with short observation series?
2 / 37
Introduction
Outline
Introduction
Markov Jump Processes
Merged Markov Jump Processes
Inference for Merged Markov Jump Processes
Generalization: Markov Switching Models
Conclusion
3 / 37
Markov Jump Processes
Finite state Markov jump process (MJP)
I Y = (Yt)t∈[0,T ) is a continuous time Markov process
I Finite state space {1, . . . , d}
I Y is time homogeneous
I Jumps of Y are governed by rate matrix Q ∈ Rd×d
I Exponential rate of leaving state k:
λk = −Qkk =∑l 6=k
Qkl < ∞
i.e. average waiting time for leaving k is 1/λk
I Conditional transition probability:
P(Yt = l |Yt− = k,Yt 6= Yt−) = Qkl/λk
4 / 37
Markov Jump Processes
Inference about rate matrix
I Ok occupation time of state k
I Nkl number of jumps from k to l
I Maximum likelihood estimation:
Qkl = Nkl/Ok
I Observing a path (Yt)t∈[0,T ), Ok and Nkl are sufficient forestimating Qkl
I Unbiased?
E(Qkl |Q) = ∞E(Qkl |Q,Ok > 0) = ∞
I Q is consistent, i.e.
limT→∞
P(|Qkl − Qkl | > ε) = 0
5 / 37
Markov Jump Processes
Some Bayesian inference
I Using uninformative prior Qkl ∼ Ga(1, 0):
Qkl |Y ∼ Ga(Nkl + 1,Ok)
I Mean: (Nkl + 1)/Ok
I Variance: (Nkl + 1)/O2k , hence decreasing with 1/T
I Mode: Nkl/Ok
6 / 37
Markov Jump Processes
Remarks on inference
I Q is well-known and looks nice
I In fact, estimation of Q is not so easy
I We need some observations for each transition k to l
I Basically, T needs to be as large as possible
7 / 37
Merged Markov Jump Processes
Introduction
Markov Jump Processes
Merged Markov Jump Processes
Inference for Merged Markov Jump Processes
Generalization: Markov Switching Models
Conclusion
8 / 37
Merged Markov Jump Processes
Observing a number of series of MJPs
I M series of MJPs are observed
I All Y (m) characterized by same rate matrix Q
I Series may be independent or come from data with breaks
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
Figure: Processes Y (1), . . . , Y (M)
9 / 37
Merged Markov Jump Processes
Observing merged MJPs
I Merged process of M single MJPs is observed
I Single MJPs characterized by same rate matrix Q
0 0.05 0.1 0.15 0.2 0.251
2
3
Figure: Merged process Y
10 / 37
Merged Markov Jump Processes
Merged MJPs
I Given: Observation process Y = (Yt)t∈[0,T )
I Y is concatenation of MJPs Y (1), . . . ,Y (M)
I In detail, Y (m) = (Y(m)t )t∈[0,T ), T = M T , and
Yt = Y(1)t if 0 ≤ t < T ,
Yt = Y(2)t−T if T ≤ t < 2T ,
...
Yt = Y(m)t−mT if (m − 1)T ≤ t < mT
I All Y (m) characterized by same rate matrix Q
I NB: Y itself is not Markov!
I Assumption of equal length is for notational convenience only
11 / 37
Merged Markov Jump Processes
Merged MJPs – Example
M = 5, d = 3, T = 0.05, T = 0.25, Q =
(−50 30 2020 −40 2030 40 −70
)
0 0.05 0.1 0.15 0.2 0.251
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
Figure: Merged process Y and single processes Y (1), . . . , Y (M)
12 / 37
Inference for Merged Markov Jump Processes
Introduction
Markov Jump Processes
Merged Markov Jump Processes
Inference for Merged Markov Jump Processes
Generalization: Markov Switching Models
Conclusion
13 / 37
Inference for Merged Markov Jump Processes
Splitting merged process
I Split Y into Y (1), . . . ,Y (M)
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
0 0.051
2
3
I Pooling:
Qkl =1
M
M∑m=1
Q(m)kl
I Problem: MLE does not exist if some O(m)k = 0;
short occupation times lead to unstable results
14 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process
I Consider Y
0 0.05 0.1 0.15 0.2 0.251
2
3
I Number of jumps and occupation time:
Nkl =M∑
m=1
N(m)kl + N+
kl , where N+kl =
M−1∑m=1
I{Y
(m)T−=k, Y
(m+1)0 =l
},
Ok =M∑
m=1
O(m)k
15 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process – first try
I First attempt: Qkl = Nkl/Ok
I Problem: Bias caused from “artificial” jumps N+kl /Ok
I If N+kl is observed explicitly, things are easy
I We assume N+kl cannot be observed
I Location of splitting points unknown
I Process not directly observed
16 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process – bias
I Assume Y(m)0 ∼ π, where π stationary distribution
I Extra jumps N+ do not affect distribution of occupation times andstationary distribution
I Hence, π = π and π = π(Q) is “unbiased” estimate for π
I Joining two independent stationary processes generates a jump fromk to l with probability πkπl :
P(N+
kl |π)
= Bin(M − 1, πkπl)
17 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process – bias (2)
I P(N+
kl |π)
= Bin(M − 1, πkπl) is justified if
I Y (m) are independent series
I Data with breaks:Length of break τ such that ρτ (Y ) is close to zero, where
ρt(Y ) =
Pdk=1 πk(Xkk(t) − πk)Pd
k=1 πk(1 − πk)
and
Xkl(t) = P(Yt = l |Y0 = k) = exp(Q t)kl
I.e. Xkk(τ) should be close to πk
18 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process – bias (3)
I Note: If π is given, N+kl and Ok are independent
I Recall: P(N+
kl |π)
= Bin(M − 1, πkπl)
I As
T
Ok
P−−−−→T→∞
1
πk,
we have
N+kl
Ok
≈ (M − 1)πkπl
πkT= πl
M − 1
T
I As π is unbiased estimate for π, these quantities can be estimatedknowing Q
19 / 37
Inference for Merged Markov Jump Processes
Estimating directly from merged process – correction
I 2-step construction for corrected estimate:
1) Qkl = Nkl/Ok
2) Qkl = Qkl − (M − 1) πl/T
I “Merging”
20 / 37
Inference for Merged Markov Jump Processes
Comparison: Splitting vs. Merging
I Variance for Splitting for Ga(1, 0) prior:
Var
(M−1
M∑m=1
Q(m)kl
∣∣∣Y (1), . . . ,Y (M)
)=
1
M2
M∑m=1
N(m)kl + 1(O
(m)k
)2I Variance for Merging for Ga(1, 0) prior:
Var(Qkl | Y
)=
∑Mm=1 N
(m)kl + 1(∑M
m=1 O(m)k
)2I For rather short single observation times T we expect Merging to
give more reliable results
21 / 37
Inference for Merged Markov Jump Processes
Numerical example
I M = 100, d = 3, T = 0.25, T = 25, Q =
(−100 60 4040 −70 3040 60 −100
)
I On average, about 22 jumps in each single process Y (m)
I Simulate merged data and apply Splitting and Merging
I Repeat 100 000 times and consider sampling distributions of Qkl
I Return process R is observable at times t = i ∆t,
Vi = ∆Ri =
∫ i ∆t
(i−1) ∆t
µs ds +
∫ i ∆t
(i−1) ∆t
σs dWs , i = 1, . . . ,H
I Vi daily stock returns
I State process Y is independent of W and not observable (hidden)
I Wanted: µ(k), σ(k), Q
I Problems:I ∆t given and fixedI Noise high compared to signalI High-frequency switching of states, i.e. λk ∆t highI Number of observations low (say, less than 5000)
31 / 37
Generalization: Markov Switching Models
Data with breaks / merged processes
I Often, we encounter data with breaks
I Weekends for daily data, nights for intra-day data, . . .
I Observable discrete time process:V = (Vi )i=1,...,H is concatenation of V (1), . . . ,V (M)
V (m) = (V(m)i )i=1,...,H and H = M H
I Hidden continuous-time state process:Y is concatenation of Y (1), . . . ,Y (M)
Y (m) = (Y(m)t )t∈[0,T ) and T = M T
32 / 37
Generalization: Markov Switching Models
Estimation from data with breaks
I Proceed similarly as for MJPs
I Use merged data V for estimation
I Employ arbitrary method to obtain estimates µ, σ, and Q
I µ and σ are not affected by merging, hence µ = µ, σ = σ
I Correction for Q: As described for MJPsI Point estimates: Correct Q to obtain Q
I Simulation based: Correct each sample Qj to obtain samples Qj
33 / 37
Generalization: Markov Switching Models
Numerical example
I T = 10, H = 10 000, ∆t = 1/1000
I d = 2, Q =(−60 60
40 −40
), µ = (2 − 1), σ = (0.10 0.05),
i.e. µ∆t = (0.002 − 0.001), σ√
∆t = (0.0032 0.0016)
I M, T , H varying s.t. T = M T and H = M H
I M = 1 corresponds to one coherent seriesM = 400 corresponds to 400 series with H = 25 observations each
I Simulate merged data
I Perform method of moments-type estimation for merged data
Table: Results for MSM (T = 10, H = 10 000): mean (top), RMSE (bottom)
35 / 37
Generalization: Markov Switching Models
Numerical example – remarks
I Estimates of µ(k), σ(k) are not affected by merging
I Quality of corrected estimates Qkl is (nearly) independent of M
I Ignoring breaks can lead to considerable bias in estimate for Q
I Method of moments-type estimation requires a lot of observations,Splitting for M > 2 not applicable;ML or Bayesian methods could be applied for M ≤ 10, but arecomputationally much more costly
36 / 37
Conclusion
Conclusion
I Estimation for set of short (independent) observation series ordata containing (long) breaks
I Applicable to processes based on MJPs
I First, estimate parameters for merged seriesSecond, correct rates for bias afterwards
I Post-processing correction
I Works with arbitrary estimation approach for coherent series
I Single series need not be of same length /splitting times need not be known –only number of breaking points required