acceptodds
Under review as a conference paper at ICLR 2027

Deep ZakaiJ: Structured Filtering for Jump-Diffusion Time Series Forecasting

Abstract

Time series driven by unobserved latent states frequently exhibit abrupt jump discontinuities whose timing and magnitude cannot be predicted from observed history alone. Classical jump-diffusion models offer a principled mathematical framework but assume rigid parametric forms, while recent neural jump models operate on fully observed trajectories without inferring the hidden states that govern the dynamics. We propose Deep ZakaiJ, a latent-state model for partially observed jump-diffusion systems that embeds the Zakai nonlinear filtering equation into a neural encoder-decoder architecture. The encoder recursively updates a belief over the latent state via Strang splitting into three interpretable substeps, namely prior propagation, diffusion innovation, and jump innovation, yielding a differentiable, first-order-accurate approximation of the exact filtering evolution. The decoder is a structured jump-diffusion model conditioned on the latent state, preserving the separation between continuous dynamics and discontinuous shocks. On synthetic and financial datasets, Deep ZakaiJ improves distributional forecasts while remaining competitive in point accuracy, achieving sharper and better-calibrated predictive intervals than the baselines and recovering interpretable latent structure in synthetic and qualitative case studies.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
latent-state estimation, nonlinear-filtering, Zakai Equation, jump-diffusion model, probabilistic time-series forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MhVdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Deep ZakaiJ, a probabilistic time series forecasting framework that combines Zakai-based latent-state filtering with a structured neural jump-diffusion decoder. The encoder uses Strang splitting to separate prior propagation, diffusion innovation, and jump innovation. Experiments on synthetic and financial datasets demonstrate improvements in probabilistic forecasting performance over various baselines.

Critical strengths
  1. The design is well-motivated. The integration of explicit latent-state filtering with jump-diffusion forecasting provides a principled and interpretable framework. Separating continuous dynamics from jump events is a reasonable and useful modeling choice.

  2. Encouraging experimental results are provided. The evaluation includes both synthetic and real-world financial datasets, with diverse baselines and multiple probabilistic metrics. The consistent improvements in CRPS and interval scores support the effectiveness of the proposed approach.

Critical weaknesses
  1. The ablation studies are conducted mainly on the synthetic dataset, and the improvements from individual filtering components are relatively small. Additional ablations on real-world datasets would better establish the necessity of the proposed structured filtering mechanism.

  2. Insufficient evaluation of future jump prediction. Although jump-aware forecasting is a central motivation, the current evaluation does not clearly distinguish predicting future jump risks from detecting regime changes after jumps have occurred. Aggregate forecasting metrics and regime-change analyses alone do not sufficiently demonstrate the model's ability to anticipate future jump events.

Questions for authors

Could the authors evaluate forecasting performance at different lead times before actual jump events, compared with periods without jumps? Event-conditioned CRPS/IS90 or jump-probability calibration would help clarify the model's ability to predict future jump risks.

1 comment
AuthorsOpenReview ↗

Response to Reviewer MhVd

Thank you for your constructive comments. We ran the paper's ablations on XAU/USD, BTC/USDT and MSTR over five seeds, and evaluated the forecast jump probability on all four datasets.

1. Ablations on the real datasets. Test segment, mean ± std over five seeds.

DatasetVariantMAERMSECRPSLogLikIS90
XAU/USDDecoder-only0.0078±0.00080.0110±0.00070.0058±0.00052.62±0.770.0557±0.0095
GRU belief0.0075±0.00000.0107±0.00000.0056±0.00012.77±0.210.0551±0.0034
Without A\mathcal A0.0086±0.00140.0139±0.00430.0065±0.00132.49±0.960.0640±0.0258
Without B\mathcal B0.0303±0.03750.0384±0.04540.0268±0.0349−5.89±8.070.4236±0.6148
Without C\mathcal C0.0076±0.00010.0107±0.00010.0055±0.00013.15±0.120.0475±0.0026
Without RψAR^{\mathcal A} _ \psi0.0080±0.00090.0120±0.00240.0059±0.00083.10±0.340.0526±0.0149
Full0.0075±0.00010.0107±0.00010.0054±0.00013.32±0.010.0433±0.0004
BTC/USDTDecoder-only0.0164±0.00700.0228±0.00820.0118±0.00422.59±0.250.0934±0.0196
GRU belief0.0105±0.00010.0160±0.00030.0080±0.00012.86±0.030.0748±0.0020
Without A\mathcal A0.0142±0.00470.0305±0.01890.0111±0.00432.65±0.350.1198±0.0737
Without B\mathcal B0.0113±0.00110.0177±0.00360.0085±0.00092.57±0.370.0788±0.0135
Without C\mathcal C0.0105±0.00010.0160±0.00020.0082±0.00012.84±0.020.0738±0.0018
Without RψAR^{\mathcal A} _ \psi0.0124±0.00180.0253±0.01060.0095±0.00152.79±0.120.0931±0.0247
Full0.0105±0.00010.0158±0.00010.0081±0.00012.87±0.010.0737±0.0014
MSTRDecoder-only0.0253±0.00600.0334±0.00550.0190±0.00281.98±0.100.1647±0.0222
GRU belief0.0170±0.00100.0256±0.00040.0130±0.00042.12±0.020.1207±0.0032
Without A\mathcal A0.0200±0.00460.0351±0.01210.0153±0.00291.91±0.180.1483±0.0194
Without B\mathcal B0.0177±0.00240.0272±0.00390.0142±0.00152.17±2.990.1807±0.0357
Without C\mathcal C0.0180±0.00320.0268±0.00280.0132±0.00172.29±0.070.1229±0.0026
Without RψAR^{\mathcal A} _ \psi0.0198±0.00540.0345±0.01590.0151±0.00402.18±0.070.1460±0.0405
Full0.0166±0.00020.0256±0.00050.0128±0.00022.29±0.060.1192±0.0030

Compared with Decoder-only, Deep ZakaiJ improves LogLik from 2.62/2.59/1.98 to 3.32/2.87/2.29 and reduces CRPS from 0.0058/0.0118/0.0190 to 0.0054/0.0081/0.0128 on XAU/BTC/MSTR, and it has higher LogLik than GRU belief on all three datasets. Removing any splitting component increases CRPS on all three datasets, and the effect can be large: without B\mathcal B, XAU/USD MAE rises from 0.0075 to 0.0303; without A\mathcal A, BTC/USDT RMSE from 0.0158 to 0.0305; without RψAR^{\mathcal A} _ \psi, MSTR RMSE from 0.0256 to 0.0345.

2. Prediction of future jumps. You asked whether the model predicts future jump risk, as distinct from detecting jumps after they occur. At forecast origin tt and lead hh, Deep ZakaiJ propagates the filtered belief through the latent dynamics and issues

p(t,h)=P(Nt+h−Nt+h−1≥1 ∣ FtX)=E[1−e−λ(θt+h−1)Δt ∣ FtX],p(t,h)=\mathbb P\bigl(N_{t+h}-N_{t+h-1}\ge1\thinspace\big|\thinspace\mathcal F_t^X\bigr)=\mathbb E\bigl[1-e^{-\lambda(\theta_{t+h-1})\Delta t}\thinspace\big|\thinspace\mathcal F_t^X\bigr],

a forecast made before the target increment is observed. NeuralMJD issues the same quantity from the jump intensity it infers from the context window. Each forecast, at every lead h∈{1,…,100}h\in\lbrace 1,\dots,100\rbrace, is scored against the true jumps on Synthetic and the 1%-level Lee–Mykland jumps on the real datasets. Higher log score and lower ECE (ten equal-count bins) are better; lead-group log scores are seed means.

DatasetModelLog score (all leads)Leads 1–50Leads 51–100ECE (×10⁻³)
SyntheticDeep ZakaiJ−0.0386±0.0000−0.0402−0.03702.934±0.624
NeuralMJD−0.0464±0.0097−0.0486−0.04424.096±1.742
XAU/USDDeep ZakaiJ−0.0190±0.0010−0.0219−0.01612.449±0.725
NeuralMJD−0.0274±0.0004−0.0320−0.02292.463±0.004
BTC/USDTDeep ZakaiJ−0.0381±0.0070−0.0443−0.03203.565±0.052
NeuralMJD−0.0419±0.0004−0.0445−0.03933.611±0.001
MSTRDeep ZakaiJ−0.0741±0.0172−0.0848−0.06357.657±0.379
NeuralMJD−0.0816±0.0008−0.0911−0.07228.147±0.005

Deep ZakaiJ has a better log score and ECE than NeuralMJD on all four datasets, and the advantage holds at leads 51–100. On Synthetic its ECE is 28% lower.

Reviewer_ocBNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies probabilistic forecasting of time series with latent states and abrupt jumps. Deep ZakaiJ combines a grid-based belief update motivated by the Zakai nonlinear filtering equation with a structured jump-diffusion decoder. The encoder uses symmetric splitting of prior propagation, diffusion likelihood updates, and jump likelihood updates. The decoder predicts drift, volatility, jump intensity, and a jump-size distribution conditional on the observation and latent state; these coefficients also define the encoder likelihood. The paper presents a first-order global error result for an idealized split update and evaluates point and probabilistic forecasts on one synthetic dataset and three financial series. Additional analyses cover component ablations, grid and architecture sensitivity, computational cost, and correspondence between latent beliefs and volatility changes. The reported advantages are strongest in CRPS and interval scores, with competitive point accuracy on the main datasets.

Critical strengths
  1. A well-motivated structured modeling contribution. The method maintains uncertainty over latent states and connects prior propagation, continuous variation, and jump updates to a shared conditional decoder. Sharing the observation model between inference and prediction gives the architecture a clear statistical motivation and a useful framework for understanding belief updates after abrupt changes.

  2. Promising probabilistic results with substantial supporting analyses. In Table 2, CRPS improves from NeuralMJD's 0.0058 to 0.0054 on XAU/USD and from 0.0085 to 0.0081 on BTC/USDT; the additional MSTR series also shows improved CRPS. The paper reports five-seed results and includes synthetic latent-state recovery, component ablations, additional baselines, and computational costs. These are positive indications of practical value, even though they do not independently establish all theoretical and calibration claims.

Critical weaknesses
  1. Specific steps in the exact-filter interpretation and error argument require clarification. A.1–A.2 provide a measure change, independence statements, and a product-rule derivation. However, observation volatility in Equation (1) depends on the hidden state; removing drift does not remove this dependence, and the derivation does not explain how Equation (3) captures state information from observed quadratic variation. Separately, A.6, line 920, equates dropping multiple-jump mixture components with capping the jump count at one, which are different approximations. Lines 922–924 then invoke normalized conditional updates to obtain an L1 bound from a prior multiple-jump probability, although the C half-step is defined as unnormalized likelihood multiplication. Theorem 5 explicitly assumes exact prior propagation, and C.7 provides empirical diagnostics, but the derivation above still requires a precise error object and stochastic interpretation.

  2. The stated sampling and interval choices need to be reconciled with the uncertainty claims. Algorithm B.3, lines 8 and 12, specifies sampling a latent state from the propagated marginal at each step without showing a conditional state transition within each trajectory. Correct latent marginals alone do not generally preserve path dependence under persistent regimes, and the impact of this approximation on multi-step forecasts is not established. Lines 374–375 explicitly define mean/standard-deviation intervals, while B.2 specifies 100 trajectories and no post-hoc standard-deviation rescaling. For non-Gaussian jump mixtures, these intervals need not be central 90% intervals of the native predictive distribution, so lower IS90 alone does not establish better calibration of that distribution. Direct coverage results and the LogLik density-estimation details remain missing.

Questions for authors
  1. Scope of the filtering equation: Equation (1) allows observation volatility to depend on the hidden state. How does Equation (3) incorporate state information from observed quadratic variation? Please specify the restrictions needed for this derivation and whether they hold for the experimental model.

  2. Multi-step prediction and calibration: Algorithm B.3 redraws the latent state from marginal beliefs at each step. How does this preserve the effect of persistent latent states on multi-step predictions? Please provide a targeted validation and report coverage of 90% predictive-quantile intervals to support the calibration claim.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ocBN

Thank you for the careful reading of the derivations.

1. Observation volatility and the derivation of Eq. (3). Proposition 1 and Eq. (3) hold for an observation volatility σ=σ(t,Xt)\sigma=\sigma(t,X_t), which is the standard setting of the Zakai equation. The drift and the jump term depend on Θt\Theta_t. The continuous part of the quadratic variation, ∫0tσ2(s,Xs)ds\int_0^t\sigma^2(s,X_s)\mathrm ds, does not depend on Θ\Theta and carries no state information. The state information enters Eq. (3) through its two innovation terms, qt(η0φ)dWtX,Qq_t(\eta_0\varphi)\mathrm dW^{X,\mathbb Q} _ t for the continuous part and qt−(η^1φ)N~X,Q(dt,dz)q_{t^-}(\hat\eta_1\varphi)\widetilde N^{X,\mathbb Q}(\mathrm dt,\mathrm dz) for the jumps. We will make this assumption explicit in Eq. (1), Proposition 1 and Appendices A.1–A.2.

The encoder applies the update of Eq. (4), a Bayes update with the increment likelihood ℓk=e−λΔt(N0+λΔtN1)\ell_k=e^{-\lambda\Delta t}(\mathcal N_0+\lambda\Delta t\mathcal N_1). This update is valid for the volatility σϕ(t,x,θ)\sigma_\phi(t,x,\theta) of the real-data model, and for σ(t,x)\sigma(t,x) its continuous-time limit is Eq. (3). On the synthetic benchmark σX\sigma_X is constant, so the assumption holds. We also trained the model with σϕ(t,x)\sigma_\phi(t,x) on the real datasets, so that the assumption holds there as well, while μ,λ,γ\mu,\lambda,\gamma and δ\delta still depend on θ\theta (test segment, mean ± std over five seeds; σϕ(t,x,θ)\sigma_\phi(t,x,\theta) is the paper's model, Tables 2 and 6):

DatasetVariantMAERMSECRPS
XAU/USDσϕ(t,x,θ)\sigma_\phi(t,x,\theta)0.0075±.00010.0107±.00010.0054±.0001
σϕ(t,x)\sigma_\phi(t,x)0.0075±.00010.0106±.00000.0054±.0001
BTC/USDTσϕ(t,x,θ)\sigma_\phi(t,x,\theta)0.0105±.00010.0158±.00010.0081±.0001
σϕ(t,x)\sigma_\phi(t,x)0.0105±.00010.0159±.00020.0082±.0001
MSTRσϕ(t,x,θ)\sigma_\phi(t,x,\theta)0.0166±.00020.0256±.00050.0128±.0002
σϕ(t,x)\sigma_\phi(t,x)0.0167±.00030.0256±.00030.0131±.0003

MAE, RMSE and CRPS agree with the paper's model to within 0.0003 on all three datasets, and CRPS stays below that of NeuralMJD (0.0058, 0.0085, 0.0137).

2. Latent sampling in Algorithm B.3. Algorithm B.3 preserves the law of the latent state at each lead; the comparison below tests the joint law of the trajectory. We compare it with a path-coherent sampler that draws θ\theta once from the filtered belief and moves it with the latent transition kernel KΔtK_{\Delta t} along each trajectory, using the energy and variogram (p=0.5p=0.5) scores of the joint 100-step trajectory and the increment CRPS of Xt+h−XtX_{t+h}-X_t, which are sensitive to path dependence.

DatasetSamplingEnergyVariogramCRPS h=1CRPS h=10CRPS h=100
SyntheticB.30.7403±.00450.02139±.000200.0032±.00000.0172±.00040.1003±.0008
Path0.7409±.00230.02138±.000220.0032±.00000.0172±.00040.1004±.0007
XAU/USDB.30.0627±.00070.00102±.000010.0009±.00000.0027±.00000.0080±.0001
Path0.0628±.00060.00102±.000010.0009±.00000.0027±.00000.0081±.0001
BTC/USDTB.30.0969±.00090.00224±.000030.0015±.00000.0034±.00000.0122±.0002
Path0.0968±.00090.00224±.000030.0015±.00000.0034±.00010.0121±.0002
MSTRB.30.1499±.00150.00341±.000150.0015±.00010.0065±.00010.0181±.0003
Path0.1493±.00100.00331±.000160.0015±.00010.0064±.00010.0180±.0002

The two samplers agree within one standard deviation over the seeds on every metric and dataset, so Algorithm B.3 reproduces the joint-trajectory scores of the path-coherent sampler. We will add this validation to Appendix B.2.

3. Quantile-interval coverage and LogLik. LogLik is the Gaussian log density log⁡N(y;μ^,σ^2)\log\mathcal N(y;\hat\mu,\hat\sigma^2) at the observation, using the same μ^\hat\mu and σ^\hat\sigma as in Section 5, without a separate density estimator. We will clarify this in Section 5. The table reports the quantile IS90 and coverage (%) of the central interval [q0.05,q0.95][q_{0.05},q_{0.95}]:

ModelSynthetic Quantile IS90Synthetic CovXAU/USD Quantile IS90XAU/USD CovBTC/USDT Quantile IS90BTC/USDT CovMSTR Quantile IS90MSTR Cov
Deep ZakaiJ0.4682±.006290.7±.30.0432±.000488.8±.40.0741±.000591.1±.20.1203±.003787.2±1.3
NeuralMJD0.7559±.333780.4±11.70.0554±.001670.4±2.10.0757±.000691.5±.70.1213±.001893.7±.7
LatentSDE1.0312±.012591.6±.30.0532±.000388.6±.30.0948±.000388.1±.10.1493±.003185.0±.4
NCDSSM0.8758±.015284.1±.40.0590±.000576.5±.50.0992±.001078.5±.50.1644±.002076.5±.6
LatentODE0.9243±.013491.4±.90.0790±.000371.0±.30.1236±.002779.7±.40.1912±.005678.8±2.3
DeepAR44.3331±64.794962.4±45.91.0360±.474826.1±36.40.1433±.019677.7±4.92.2199±2.411450.9±35.7
Chronos1.5106±.014855.3±.20.0733±.001166.5±.40.1040±.000874.2±.30.1891±.001262.9±.4

Deep ZakaiJ achieves the lowest quantile IS90 on all four datasets, with coverage close to the nominal 90%.

4. Error object in the proof of Theorem 5. The at-most-one-jump likelihood ℓk\ell_k drops the components n≥2n\ge2 of the exact increment likelihood. For each θ\theta,

rk(θ):=ℓkexact(θ)−ℓk(θ)=∑n≥2e−λΔt(λΔt)nn!N(ΔXk;μΔt+nγ,σ2Δt+nδ2)≥0.r_k(\theta):=\ell_k^{\mathrm{exact}}(\theta)-\ell_k(\theta)=\sum_{n\ge2}e^{-\lambda\Delta t}\frac{(\lambda\Delta t)^n}{n!}\mathcal N\bigl(\Delta X_k;\mu\Delta t+n\gamma,\sigma^2\Delta t+n\delta^2\bigr)\ge0 .

Part (II) of the proof in A.6 compares, for a normalized density qq, the unnormalized update S~Δt(q)=ℓk1/2⊙SΔtA(ℓk1/2⊙q)\widetilde{\mathcal S} _ {\Delta t}(q)=\ell_k^{1/2}\odot S^{\mathcal A} _ {\Delta t}(\ell_k^{1/2}\odot q) with the same update SΔtStrang(q)\mathcal S^{\mathrm{Strang}} _ {\Delta t}(q) built from ℓkexact=ℓk+rk\ell_k^{\mathrm{exact}}=\ell_k+r_k, where SΔtAS^{\mathcal A} _ {\Delta t} has transition density pΔt(θ,θ′)p_{\Delta t}(\theta,\theta'). Both are unnormalized joint densities of the next latent state and the increment ΔXk\Delta X_k; we bound their difference in L1(dθ⊗dΔXk)L^1(\mathrm d\theta\otimes\mathrm d\Delta X_k). Since ℓkexact≥ℓk≥0\ell_k^{\mathrm{exact}}\ge\ell_k\ge0, the difference is nonnegative, and Fubini gives

\qquad \rho(\theta,\theta')=\int\sqrt{\ell_k(\theta)\ell_k(\theta')}\mathrm d\Delta X_k .$$ The identity $\sqrt{ab}=\tfrac12(a+b)-\tfrac12(\sqrt a-\sqrt b)^2$ splits the bracket into the dropped mass and a Hellinger correction, $$\rho^{\mathrm{exact}}-\rho=\tfrac12\bigl[\mathbb P_\theta(N_{\Delta t}\ge2)+\mathbb P_{\theta'}(N_{\Delta t}\ge2)\bigr]-\bigl(H^2_{\mathrm{exact}}-H^2\bigr)(\theta,\theta'),\qquad H^2(\theta,\theta')=\tfrac12\bigl\Vert\sqrt{\ell_k(\theta)}-\sqrt{\ell_k(\theta')}\bigr\Vert_{L^2}^2 ,$$ where $\int r_k(\theta)\mathrm d\Delta X_k=\mathbb P_\theta(N_{\Delta t}\ge2)\le\tfrac12(\Lambda_\varepsilon\Delta t)^2$. This prior probability appears because the error object integrates over the increment. With $d_\theta=\sqrt{\ell^{\mathrm{exact}} _ k(\theta)}-\sqrt{\ell_k(\theta)}$, expanding the square gives $$\bigl|H^2_{\mathrm{exact}}-H^2\bigr|\le\bigl\Vert\sqrt{\ell_k(\theta)}-\sqrt{\ell_k(\theta')}\bigr\Vert_{L^2}\Vert d_\theta-d_{\theta'}\Vert_{L^2}+\tfrac12\Vert d_\theta-d_{\theta'}\Vert_{L^2}^2 .$$ Assume decoder coefficients bounded and Lipschitz in $\theta$, $\sigma(t,x)\ge\underline\sigma > 0$, and Gaussian jump sizes with $\delta\ge\underline\delta > 0$ or a $\theta$-independent constant jump size (synthetic). Since $$\partial_\theta d_\theta=\frac{\partial_\theta r_k}{2\sqrt{\ell_k^{\mathrm{exact}}}}-\frac{\partial_\theta\ell_k}{2\sqrt{\ell_k}}\cdot\frac{d_\theta}{\sqrt{\ell_k^{\mathrm{exact}}}}, \qquad 0\le d_\theta\le\min\Bigl(\sqrt{r_k},\frac{r_k}{2\sqrt{\ell_k}}\Bigr),$$ and each component of $r_k$ has weight $(\lambda\Delta t)^n$ with $n\ge2$, both terms are $O(\Delta t)$ in $L^2$: $\Vert\partial_\theta\sqrt{r_k}\Vert_{L^2}\le C\Delta t$ and $\int(\partial_\theta\sqrt{\ell_k})^2\min(1,r_k/\ell_k)\mathrm d\Delta X_k\le C\Delta t^2$. Hence $\Vert d_\theta-d_{\theta'}\Vert_{L^2}\le L_d\Delta t|\theta-\theta'|$, and $\Vert\sqrt{\ell_k(\theta)}-\sqrt{\ell_k(\theta')}\Vert_{L^2}\le L_H\bigl(|\theta-\theta'|+(\Delta t|\theta-\theta'|)^{1/2}\bigr)$, where the second term covers latent values at which $\lambda$ vanishes. Using $\int p_{\Delta t}(\theta,\theta')|\theta-\theta'|^2\mathrm d\theta\le c_K\Delta t$, and $\int q=1$, for $\Delta t\le1$, $$\mathrm{(II)}\le\tfrac12(\Lambda_\varepsilon\Delta t)^2+C'\Delta t^2,\qquad C'=c_KL_d(L_H+L_d)+c_K^{3/4}L_HL_d ,$$ which has the form $C_1\Delta t^2+C_2(\Lambda_\varepsilon\Delta t)^2$ of Theorem 5. For the global bound, each update is a positive linear map with $\rho^{\mathrm{exact}}\le1$ and $\int p_{\Delta t}(\theta,\theta')\mathrm d\theta=1$, so its output has at most the mass of its input and it is a contraction in $L^1(\mathrm d\theta\otimes\mathrm d\Delta X_k)$. Telescoping over $n=T/\Delta t$ steps accumulates (II) to $T(\tfrac12\Lambda_\varepsilon^2+C')\Delta t$ in $L^1(\mathrm d\theta\otimes\mathrm d\Delta X_{1:n})$, first order as in Corollary 6. For the normalized belief, let $q_n$ be the exact unnormalized filter after $n$ steps and $\tilde q_n=\widetilde{\mathcal S} _ {\Delta t}^{n}(q)$, with normalized beliefs $\pi_n$ and $\tilde\pi_n$. For each observation path, Proposition 7 gives $\Vert\tilde\pi_n-\pi_n\Vert_1\le2\Vert\tilde q_n-q_n\Vert_{L^1(\mathrm d\theta)}/\Vert q_n\Vert_{L^1(\mathrm d\theta)}$, and $\Vert q_n\Vert_{L^1(\mathrm d\theta)}$ is the density of the observed increments $\Delta X_{1:n}$. Integrating over $\Delta X_{1:n}$ gives $$\mathbb E\Vert\tilde\pi_n-\pi_n\Vert_1\le2\Vert\tilde q_n-q_n\Vert_{L^1(\mathrm d\theta\otimes\mathrm d\Delta X_{1:n})},$$ with the expectation over the observation path. Hence a bound in this norm bounds the expected $L^1$ error of the normalized filter. We will replace part (II) in Appendix A.6 by this argument, stating its error object and conditions.
Reviewer_FM2EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Deep ZakaiJ, a latent-state model for partially observed jump-diffusion systems that embeds the Zakai nonlinear filtering equation into a neural encoder–decoder architecture. The model was evaluated by one synthetic dataset and two real world financial data. The results showed that the model performs strongly among the evaluated baselines.

Critical strengths
  1. Lots of experiments with different datasets have been done to evaluate the proposed framework and compared with the existing work.
  2. The model achieves strong probabilistic forecasting performance.
Critical weaknesses
  1. Neural approximations and splitting schemes for the Zakai equation have already been studied in prior work, such as Bågmark et al. (2023). The main contribution seems to be integrating existing techniques into an end-to-end forecasting framework. Technical innovation is not strong.
  2. There is no jump detection accuracy metric.
Questions for authors
  1. All reported metrics are aggregate and dominated by non-jump periods. Can you report the jump detection accuracy for both synthetic and real world data?
  2. Given the substantial computational overhead, what practical advantages does Deep ZakaiJ offer over simpler neural jump-diffusion models such as NeuralMJD?
  3. Does the first-order approximation guarantee still hold when the learned residual correction is included in the encoder?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.