Pith. sign in

REVIEW 3 major objections 6 minor 16 references

Hybrid SIS Dynamics for Demand Modeling of Frequently Updated Products

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proves that a hybrid SIS demand model with instantaneous post-update jumps has uniquely identifiable parameters whenever each update interval contains two distinct state observations.

desk verdict The identifiability theorem is solid and citable; the real-data validation is an in-sample fit of a mismatched state variable, so the empirical half needs major revision. read the letter →

arxiv 2506.01866 v1 pith:W627XT3L submitted 2025-06-02 eess.SY cs.SY

Ian Walter , Jitesh H. Panchal , Philip E. Paré This is my paper
classification eess.SYcs.SY MSC 93B3092D30
keywords hybridsystemmodelSISepidemicdynamicsparameteridentifiabilitydemandmodelinglive-servicevideogamesimpulseeventsEulerdiscretizationreproductionnumber
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Live-service software products such as online games tend to show sudden demand spikes immediately after an update, followed by slower engagement dynamics until the next release. The authors model this as a hybrid spreading process: between updates, the fraction of active users follows continuous-time susceptible-infected-susceptible (SIS) dynamics, and at each update the state jumps by a multiplicative factor. Their central result is a necessary-and-sufficient condition for identifying every model parameter from sampled data: each interval must contain at least two distinct demand observations, the interval must be long enough, and the demand just before an update must be nonzero. When these conditions hold, parameter estimation is a single linear solve, and the paper shows the resulting model tracks real daily player counts for the live-service game Apex Legends across five update intervals.

What carries the argument

The carrying object is the block-diagonal observation matrix $\Psi = \mathrm{diag}(\Phi_0, \ldots, \Phi_m)$ built from the observed state differences. Each interior block $\Phi_i$ stacks one column for the jump parameter $\alpha_i$, whose entry is the pre-update state $x_{T_i-1}$, and two columns for the interval's SIS parameters $\beta_i$ and $\gamma_i$, whose entries are the Euler-discretized terms $h(1-x_k)x_k$ and $-h x_k$. The theorem's conditions are exactly the conditions under which every block has full column rank, making the linear system (10) uniquely solvable.

What would settle it

Apply the estimator to a real or simulated demand series in which three consecutive daily state values are exactly equal within one update interval; Theorem 1 predicts the interval parameters cannot be uniquely identified, so if the routine returns a unique parameter vector with zero residual, the claimed necessary condition is wrong.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1: for the discretized hybrid SIS model in (2), under the assumptions that update times and all state values are known, the full parameter vector $\Theta = [\beta_0, \gamma_0, \alpha_1, \ldots, \alpha_m, \beta_m, \gamma_m]'$ is uniquely identifiable if and only if the first interval has more than two steps and two observations satisfying (3a), every intermediate interval has more than two steps, two such observations, and a nonzero pre-update state, and the final interval has more than one step plus the same two-observation and nonzero-state conditions. The proof rewrites the state differences as a block-diagonal linear system $\Psi\Theta = \text{state differences}$ and observes that unique identifiability is exactly the statement that $\Psi$ has full column rank. Conditions (3a)--(3c) are precisely the rank conditions for each block.

Load-bearing premise

The whole identification argument assumes the update times and the demand value at every sampled time step are known exactly; in the real-data validation this means the manually selected 'significant update' dates must be the true impulse times, since each jump size is read off from the single state difference across that one date.

Editorial extensions

If this is right

  • If the rank conditions are met, all jump sizes and interval spread/decay rates are recovered in one linear solve, with no iterative fitting or prior parameter guesses.
  • The reproduction number $R_0 = \beta/\gamma$ per interval stays close to its true value under both observation and process noise (within 8% and 4%, respectively, in the simulations), so equilibrium demand levels are the most trustworthy outputs.
  • Intervals shorter than two steps, or intervals where the state is constant across all observations, cannot identify the parameters; a practitioner must ensure each interval contains at least two distinct observations.
  • On real daily player-count data, the fitted hybrid model captures the overall demand trajectory across five update intervals, although the simulated spikes over- or undershoot the actual spikes in later intervals.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the theorem assumes update times are known, an automatic update-date detector would turn the method into an online forecasting tool; that extension changes the problem, since inferred jump times would themselves be parameters.
  • Each jump size $\alpha_i$ is estimated from the single state difference across one chosen date, so routine daily fluctuations (weekday/weekend effects) can bias it; using a short window of post-update observations could stabilize the estimate but would require relaxing the instantaneous-impulse assumption.
  • The same block-rank argument should carry over to other one-dimensional spreading models with impulsive resets (logistic growth, SIR without demography), because only the two columns representing the continuous dynamics change.
  • A natural testable extension is to replace the manually selected update dates with official patch-release dates for Apex Legends and check whether the fitted jump sizes align with the actual content changes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a hybrid continuous-time SIS model in which a demand state evolves according to SIS dynamics between product updates, with an instantaneous multiplicative jump at each update. A Euler-discretized version is introduced, and the paper proves necessary and sufficient conditions for unique identification of the full parameter vector when update times and state values are known at every discrete step. The identifiability theorem is verified in simulation for noiseless, observation-noisy, and process-noisy synthetic data generated from the continuous-time model, and the method is then applied to daily peak concurrent player counts for Apex Legends. The theoretical result appears internally consistent; the main weakness is the real-data validation, where the daily-peak measurement is not the instantaneous state assumed by the model and the update dates are selected without a stated rule.

Significance. The identifiability theorem is the paper's main contribution and is a useful, clean result: it gives explicit rank conditions under which all hybrid model parameters can be recovered by solving a linear system, and the proof is self-contained. The simulation study is also a genuine strength: it tests the estimator against known ground truth, reports error as a function of step size, and compares the reproduction number rather than only raw parameters. If the result holds, it provides practitioners with a practical criterion for when their observation schedule is sufficient for parameter recovery. However, the paper's claim that the model is validated on real user engagement data is not supported as written, because the dataset consists of daily maxima while the model state is an instantaneous proportion; this is a structural data-model mismatch rather than a mere implementation detail.

major comments (3)
  1. [Section V.B, Section IV, Eq. (1a)-(2b)] The real-data validation uses daily peak concurrent player counts, as stated in Section IV, but the model state x(t) in Eq. (1a) is the instantaneous proportion of active users and the estimator (2) assumes point samples x_k of that state. The daily maximum is an order statistic of the within-day trajectory, so it does not evolve according to (2a), and the impulse relation (2b) between the peak on the update day and the peak on the previous day equals the jump factor only under unstated conditions about when the update occurs and when the peaks occur relative to the jump. Since Section V.A validates the estimator only on point samples of the continuous-time model, this interface is never tested. The statement in Section V.B that the model 'can capture real demand behavior' is therefore not established by Figure 5; the fit is an in-sample fit of a different, unstated discrete-time model whose state is the daily peak. The validation claim should be reframed, or the model should be augmented with an observation equation linking the daily maximum to the continuous-time trajectory.
  2. [Section V.B, Assumption 1, Eq. (2b)] The number of updates m and the update times T_i are taken as known and are shown as dashed vertical lines in Figure 2, but the manuscript never specifies how these 'significant update' dates were selected or how robust the estimates are to mis-specification. Since each α_i is estimated from exactly one state difference across the selected date in Eq. (2b), a misassigned or missed update would corrupt the block-diagonal structure in (10) and bias the subsequent β_i and γ_i estimates. The authors should state the selection rule (patch notes, changelog, or a threshold criterion) and provide a sensitivity analysis with respect to perturbed T_i values.
  3. [Section V.B] No quantitative error metric is reported for the real-data fit. The text acknowledges that simulated spikes over- or undershoot the real data and that later intervals reach equilibrium too quickly, but 'good fit' is supported only by visual inspection. To make the validation claim meaningful, the paper should report per-interval normalized RMSE or a similar metric, compare against a baseline model, and clearly separate in-sample fit from out-of-sample prediction; the current prediction discussion in Section V.B remains qualitative.
minor comments (6)
  1. [Section II, notation] The notation 'I ∈ {0} S[N]' appears to be a typesetting error; it should presumably read I ∈ {0,1,...,N} or equivalent set-builder notation.
  2. [Section III, Eq. (10)] The paper uses both a prime and a superscript T for transpose; for consistency, one notation should be used throughout.
  3. [Section V.A.2] The sensitivity example with β1=1.9 and γ1=1.5 is described in the text but not shown in any figure; either add the corresponding plot or present the result as a supplementary experiment.
  4. [Section V.A.3] The process-noise simulation using Eq. (15) does not specify the numerical integration scheme or time step used for the SDE; adding this detail would improve reproducibility.
  5. [Figure 2] The caption states that dashed lines represent significant update dates but gives no source or criterion for identifying them; at minimum, the caption should refer to the procedure described in the text.
  6. [References] Reference [14, Thm 7] is cited for a rank condition for the SIS model; since [14] is an overview article, a more precise citation to a theorem or proposition would be helpful.

Circularity Check

1 steps flagged · score 5.0 of 10

Theorem 1 is a self-contained rank argument, but the real-data 'validation' is an in-sample fit: the update-jump parameters are estimated from the very observations the simulation is then said to reproduce, a point the paper itself concedes.

  1. fitted input called prediction [Section V.B (Model Validation with Real Data), Figure 5]
    "We estimate the parameters for multiple updates simultaneously, assuming a total population size of N = 1,000,000, and use these parameters to generate a fit for the data using our hybrid model. ... Overall, the simulation provided a good fit for the data, and instills confidence that the model in (1a)–(1b) can capture real demand behavior. However, this fit assumes all data was known, while a dynamic demand model would likely be used in an online fashion to predict future demand."

    The parameters are solved from (10), which is constructed from all observed x_k under Assumption 2; the 'validation' simulation is then generated from those same fitted parameters plus the observed initial value. In particular, α_i is algebraically determined by the two observations around each update through (2b), so the simulated update spikes are forced to match the data rather than independently predicted. The paper's own statement that the fit assumes all data was known confirms this is an in-sample fit. Thus the claim that the model captures real demand behavior is supported by a restatement of the fitted inputs, not by an out-of-sample check.

full rationale

The identifiability theorem and its proof (Section III) are not circular: given Assumptions 1 and 2, the discrete model (2) is rewritten as the linear system y = ΨΘ, and full column rank of Ψ is shown equivalent to the stated interval-length and distinct-state conditions. This is a legitimate rank argument, and the simulation studies in Section V.A compare estimates to known ground truth, which is standard estimator validation rather than circularity. The only overlapping-author citation, [14, Thm 7], is used for the first interval's rank condition; it is not load-bearing in a circular sense because the same elementary rank reasoning is proved in Lemma 1 for the other blocks and concerns plain SIS identifiability, not the hybrid model. The real-data section, however, validates the model by fitting all parameters to the same daily data and then forward-simulating from those fitted parameters, with the update magnitudes α_i pinned by the two observations around each update via (2b). The paper explicitly concedes 'this fit assumes all data was known.' That makes the real-data 'good fit' partly circular as a validation claim, even though it does not affect the mathematical identifiability result. A separate data-model mismatch (daily peak concurrent users vs. instantaneous state x_k) is a correctness concern, not a circularity, and is not scored here.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The core model parameters α, β, γ are fitted to the same real dataset used for the claimed validation, which raises circularity. The population size N and the update times are chosen by hand, and the latter are not justified. No new physical entities are introduced; the impulse factor is a model parameter, not an invented entity.

free parameters (5)
  • Population size N = 1,000,000
    Chosen by hand in Section V.B to convert daily peak player counts to fractions x = I/N. The fitted β, γ, α depend on this scaling and no sensitivity study is provided.
  • Update jump factors α_i (i=1..4) = estimated from Apex data (values shown at top of Figure 5)
    Fitted to real data via the linear system (10); these are the impulse magnitudes.
  • Engagement rates β_i and disinterest rates γ_i = estimated from Apex data (shown at top of Figure 5)
    Fitted to real data; the paper claims these capture demand dynamics but the fit is in-sample.
  • Noise strengths σ = 0.02 (observation and process noise)
    Hand-chosen for the simulation studies in Section V.A; not varied or justified.
  • Update time indices T_1,...,T_m = selected manually from SteamDB chart
    The paper does not state selection criteria; estimates are sensitive to these choices (Section V.B).
assumptions (4)
  • domain assumption The demand process is well described by SIS dynamics with homogeneous mixing (1a).
    The model assumes users can freely start/stop and engagement spreads like an epidemic in a well-mixed population; no network structure or heterogeneity is included (Section II).
  • domain assumption The Euler discretization (2a) with daily time step h=1 adequately approximates the continuous-time dynamics for the real data.
    The estimation and real-data fit use the discrete model (2); the paper's own simulations show discretization error at h=1 up to 1.25% for the reproduction number in the noiseless case (Section V.A.1).
  • domain assumption All state values x_k and step size h are exactly known (Assumption 2).
    The identifiability theorem requires perfect observation of the state; the real data is noisy, and the noise analysis only quantifies estimation error, not identifiability.
  • standard math Reference [14, Theorem 7] correctly characterizes identifiability of the standard discrete-time SIS parameters.
    Used in the proof of Theorem 1 to establish the first-interval conditions; this is a cited prior result by one of the authors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid SIS Dynamics for Demand Modeling of Frequently Updated Products." pith.science (2026). https://pith.science/paper/W627XT3L

@misc{pith2026250601866,
  author       = {Pith},
  title        = {Pith review of: Hybrid SIS Dynamics for Demand Modeling of Frequently Updated Products},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W627XT3L}},
  note         = {Machine review of arXiv:2506.01866}
}
read the original abstract

We propose a hybrid spreading process model to capture the dynamics of demand for software-based products. We introduce discontinuous jumps in the state to model sudden surges in demand that can be seen immediately after a product update is released. After each update, the modeled demand evolves according to a continuous-time susceptible-infected-susceptible (SIS) epidemic model. We identify the necessary and sufficient conditions for estimating the hybrid model's parameters for an arbitrary finite number of sequential updates. We verify the parameter estimation conditions in simulation, and evaluate how the estimation of these parameters is impacted by the presence of observation and process noise. We then validate our model by applying our estimation method to daily user engagement data for a regularly updating software product, the live-service video game `Apex Legends.'

Figures

Figures reproduced from arXiv: 2506.01866 by the authors.

Figure 1
Figure 1. Example system graph for a system with two update [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Player count dataset. The plot shows the peak number [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The simulated data from (1a)–(1b) used for validating [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The top row of plots shows the relative error of the estimated model parameters from the true parameters used [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Real data and simulation results. The blue line is [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 15 canonical work pages

  1. [1]

    The origins and development of the diffusion of innovations paradigm as an example of scientific growth,

    T. W. Valente and E. M. Rogers, “The origins and development of the diffusion of innovations paradigm as an example of scientific growth,” Science Communication , vol. 16, no. 3, pp. 242–273, 1995. [Online]. Available: https://doi.org/10.1177/1075547095016003002

    work page doi:10.1177/1075547095016003002 1995

  2. [2]

    E. M. Rogers, Diffusion of Innovations . Simon and Schuster, 2010

    work page 2010

  3. [3]

    The spread of innovations in social networks,

    A. Montanari and A. Saberi, “The spread of innovations in social networks,” Proceedings of the National Academy of Sciences , vol. 107, no. 47, pp. 20 196–20 201, 2010. [Online]. Available: https://www.pnas.org/doi/abs/10.1073/pnas.1004098107

    doi:10.1073/pnas.1004098107 2010

  4. [4]

    Going viral: Stability of consensus-driven adoptive spread,

    S. F. Ruf, K. Paarporn, and P. E. Par ´e, “Going viral: Stability of consensus-driven adoptive spread,” IEEE Transactions on Network Science and Engineering , vol. 7, no. 3, pp. 1764–1773, 2020

    work page 2020

  5. [5]

    Diffusion dynamics in small-world networks with heterogeneous consumers,

    S. A. Delre, W. Jager, and M. A. Janssen, “Diffusion dynamics in small-world networks with heterogeneous consumers,” Computational and Mathematical Organization Theory , vol. 13, pp. 185–202, 2006

    work page 2006

  6. [6]

    Agent- based simulation of innovation diffusion: A review,

    E. Kiesling, M. G ¨unther, C. Stummer, and L. Wakolbinger, “Agent- based simulation of innovation diffusion: A review,”Central European Journal of Operations Research , vol. 20, pp. 183–230, 2011

    work page 2011

  7. [7]

    Virus spread in networks,

    P. Van Mieghem, J. Omic, and R. Kooij, “Virus spread in networks,” IEEE/ACM Transactions on Networking, vol. 17, no. 1, pp. 1–14, 2009

    work page 2009

  8. [8]

    Diffusion in social networks as SIS epidemics: Beyond full mixing and complete graphs,

    J. Zhang and J. M. F. Moura, “Diffusion in social networks as SIS epidemics: Beyond full mixing and complete graphs,” IEEE Journal of Selected Topics in Signal Processing , vol. 8, no. 4, pp. 537–551, 2014

    work page 2014

Show all 16 references
  1. [9]

    De Tarde, The Laws of Imitation

    G. De Tarde, The Laws of Imitation . H. Holt, 1903

    1903

  2. [10]

    On a networked SIS epidemic model with cooperative and antagonistic opinion dynamics,

    B. She, J. Liu, S. Sundaram, and P. E. Par ´e, “On a networked SIS epidemic model with cooperative and antagonistic opinion dynamics,” IEEE Transactions on Control of Network Systems , vol. 9, no. 3, pp. 1154–1165, 2022

    2022

  3. [11]

    Modelling a dynamic market potential: A class of automata networks for diffusion of innovations,

    R. Guseo and M. Guidolin, “Modelling a dynamic market potential: A class of automata networks for diffusion of innovations,” Technological F orecasting and Social Change , vol. 76, no. 6, pp. 806–820, 2009. [Online]. Available: https://www.sciencedirect.com/ science/article/pi...

    2009

  4. [12]

    Extending LaSalle’s invariance principle to impulsive switched systems with an application to hybrid epidemic dynamics,

    J. Liu, X. Liu, and W.-C. Xie, “Extending LaSalle’s invariance principle to impulsive switched systems with an application to hybrid epidemic dynamics,” in Proceedings of the Chinese Control and Decision Conference, 2010, pp. 136–141

    2010

  5. [13]

    Atkinson, An Introduction to Numerical Analysis

    K. Atkinson, An Introduction to Numerical Analysis . John wiley & sons, 1991

    1991

  6. [14]

    Modeling, estimation, and analysis of epidemics over networks: An overview,

    P. E. Par ´e, C. L. Beck, and T. Bas ¸ar, “Modeling, estimation, and analysis of epidemics over networks: An overview,” Annual Reviews in Control , vol. 50, pp. 345–360, 2020. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S1367578820300614

    2020

  7. [15]

    Pay for play – behavioural patterns of pay-to-win gaming,

    B. Lelonek-Kuleta, R. P. Bartczuk, and M. Wiechetek, “Pay for play – behavioural patterns of pay-to-win gaming,” Computers in Human Behavior, vol. 115, p. 106592, 2021

    2021

  8. [16]

    Analysis of an innovative business model: The freemium and its applications,

    L. De Picquendaele and M. M. Duarte, “Analysis of an innovative business model: The freemium and its applications,” Master’s thesis, Nova SBE, Carcavelos, Portugal, 2016

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.