Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Forecasting Russian Equipment Losses Using Time Series and Deep Learning Models

T0 review · 5 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that TCN and LSTM networks give the steadiest forecasts of Russian equipment losses, with all five models projecting a plateau or decline into late 2025.

desk verdict A naive comparison of five forecasting models on a single OSINT loss series; honest about data limitations but missing any quantitative evaluation, so its central claims are unsupported. read the letter →

arxiv 2509.07813 v1 pith:27CHHKPP submitted 2025-09-09 cs.LG cs.AI

Jonathan Teagan This is my paper · ORCID
classification cs.LGcs.AI
keywords Russia-UkrainewarmilitaryforecastingOSINTdatamateriallossesLSTMtemporalconvolutionalnetworksensembletimeseriesmodels
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

This paper tries to show that standard time-series and deep-learning forecasting tools can make useful short-term predictions from open-source tallies of Russian military equipment destroyed in Ukraine. It fits five models—ARIMA, Prophet, LSTM, TCN, and XGBoost—to daily and monthly counts of visually confirmed losses and argues that the deep-learning models, especially TCN and LSTM, produce the steadiest and most consistent forecasts at daily resolution. The author also claims that all five models, despite their differences, agree on a broad near-term trend: Russian equipment losses are stabilized or declining through late 2025, with only XGBoost forecasting a modest rise. The reason this matters is that no reliable state-issued loss numbers exist, so a method that can extract a signal from noisy, crowd-sourced visual confirmations would give analysts, policymakers, and researchers a low-cost quantitative window into material attrition. The whole argument hinges on one assumption: that the fluctuations in the confirmed-loss series reflect battlefield destruction rather than reporting delays and platform moderation.

What carries the argument

The load-bearing object is a cleaned, geocoded time series of visually confirmed Russian equipment destructions—tanks, IFVs, and support vehicles—each entry dated and grouped by district and province, deduplicated, with a deliberately incomplete month excluded from training. The forecasting machinery consists of five comparators: ARIMA as a linear statistical baseline, Prophet for additive trend and seasonality decomposition, LSTM with gated memory cells for long-range sequence dependencies and capturing weekday confirmation rhythms, TCN with dilated causal convolutions (convolutions that skip steps to see both near and distant context) for parallel multi-scale sequence modeling, and XGBoost

What would settle it

Track an independent, non-visual measure of Russian armor destruction for the same June-December 2025 window—satellite-detected wreckage, unit-level loss reports, or inventory surveys—against the OSINT series. If the independent series rises while the OSINT counts and forecasts stay flat, or if a documented social-media moderation crackdown produces an immediate drop across all equipment categories, the plateau is an artifact of confirmation flow rather than attrition.

Watch

Extended reading notes

Core claim

The central claim is a comparative empirical one: among the five architectures tested, TCN and LSTM yield the most stable and internally consistent forecasts of Russian equipment losses, particularly when the data are modeled at daily rather than monthly granularity. Running the same forecasting problem through classical and deep-learning models and comparing their outputs, the paper reports a cross-model consensus—TCN and LSTM project a plateau, Prophet projects a slow decline, and XGBoost anticipates a modest second-half-2025 uptick—which the author reads as evidence that the near-term trajectory is bounded even though short-term escalation cannot be ruled out. The paper further argues tha

Load-bearing premise

The forecasts stand on the assumption that the daily counts of visually confirmed equipment losses vary with actual battlefield attrition—not with how quickly and how completely destroyed vehicles get photographed, uploaded, and approved; the paper's own limitations section says that assumption is not fully met.

Editorial extensions

If this is right

  • If TCN and LSTM forecasts are reliable, sudden departures from their predicted plateau become a practical early-warning signal for escalation or for breaks in OSINT reporting.
  • If all five models are right, Russian armored vehicle losses will stay near current levels through late 2025 rather than rising sharply, constraining expectations about near-term force regeneration.
  • The models' disagreement—XGBoost's uptick versus the others' plateau—shows that single-model forecasts on this kind of data are insufficient, and ensemble interpretation is the defensible use.
  • Daily resolution appears to be where deep-learning models add the most value, which argues for maintaining high-granularity OSINT collection rather than aggregating to monthly counts.

Reading between the lines

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

  • Because the paper itself documents confirmation lags of weeks and strong weekday periodicity in the counts, much of what LSTM learns as 'seasonality' is likely a documentation rhythm; a model that separates confirmation delay from combat occurrence would test this directly.
  • The geocoded regional fields are never used as predictors; using them could reveal whether forecasts are dominated by the few well-documented districts, and whether the plateau is a media-coverage artifact.
  • An independent check against non-visual loss indicators, such as satellite-detected vehicle destruction or inventory estimates, would settle whether the plateau is real attrition or reduced visibility.
  • The method could transfer to other conflicts using similarly constructed OSINT tabulations, with the same caveat: the forecast is only as meaningful as the visibility process behind the counts.
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

5 major / 4 minor

Summary. The paper applies ARIMA, Prophet, LSTM, TCN, and XGBoost to WarSpotting OSINT counts of Russian equipment losses (February 2022–July 2025) and presents forecasts through late 2025. The central claims are that TCN and LSTM produce stable and consistent forecasts, and that all models agree on a plateau or decline in Russian equipment loss rates. Data preprocessing excludes June 2025 (and in places June–July 2025) due to incomplete reporting. The paper provides no quantitative accuracy evaluation, no hyperparameter details, no reproducible code or data, and no correction for the OSINT biases it itself enumerates.

Significance. If the central claim were supported, the paper would offer a useful comparison of classical and deep learning forecasting methods on a novel OSINT-derived conflict dataset, with practical implications for open-source attrition monitoring. The breadth of models and the explicit acknowledgment of OSINT limitations are strengths. However, the paper as it stands does not validate the models: there is no holdout evaluation, no error metric, no uncertainty quantification for most models, and the forecast values contradict each other. Moreover, the paper's own limitations section lists mechanisms—confirmation lags, undercounting, spatial skew, platform moderation—that make the WarSpotting series a plausible proxy for OSINT visibility rather than battlefield attrition. The central claims about TCN/LSTM stability and an agreed plateau are therefore unsupported at the evidentiary and construct-validity levels.

major comments (5)
  1. [§2.2 and 'Model Interpretations and Forecast Analysis' (Figs. 1–11)] No holdout evaluation or error metrics are reported anywhere. The claims that TCN and LSTM are 'stable and consistent' and that 'all models agree' are based on visual inspection of in-sample fits and free-running forecasts, with no RMSE/MAE, no backtest, no confidence intervals for most models, and no comparison on a common forecast origin. This is load-bearing because the abstract and conclusion rest on these stability claims, and the current evidence cannot distinguish model skill from curve fitting.
  2. [ARIMA Forecasts; Figure 6] The reported forecast values are internally contradictory. The ARIMA text says 'projected monthly attrition remaining near 80–100 vehicles' but then describes the ARIMA forecast as 'a modest rebound and stabilization of losses around 63–65 units per month.' Similarly, the XGBoost tank-only forecast is described as '3–5 vehicles per day' (text) and '4–8 losses per day' (Figure 6 caption). These are not minor wording differences; they change the predicted attrition level by ~50%. No explanation is offered, and the reader cannot tell which number the author considers the model output.
  3. [§2.1 and 'ARIMA Forecasts (Excluding June–July 2025)'] The exclusion of incomplete months is inconsistent. Methodology (§2.1) says June 2025 was excluded; the ARIMA section says 'the months of June and July 2025 were excluded' and that the model was trained on 'February 2022 to May 2025 and August 2025 onward', although the dataset ends in early July 2025. Figure 1 is described as covering 'January 2022 through July 2025' while the text says June–July are excluded. Since the excluded data directly affect the forecast trend, this instability in the data pipeline undermines the reported forecasts.
  4. [Limitations of OSINT-Based Loss Data (WarSpotting)] The paper's own limitations section identifies structural undercounting, confirmation lags of weeks, regional/platform sampling bias, and platform moderation as features of the WarSpotting data. No measurement-error model, external validation, or benchmark against independent loss estimates is provided. Given these mechanisms, a decline in recorded counts can arise from reduced documentation even if losses are unchanged. The conclusion that 'Russian equipment loss rates' are stabilized or declining is therefore not supported; the forecasts may be tracking OSINT visibility rather than attrition. This is a construct-validity problem, not a minor caveat.
  5. [§2.2 and 'Conclusion'] The paper claims model superiority ('TCN and LSTM offered the most temporally consistent forecasts') without reporting any quantitative comparison, training details, or hyperparameter settings. No train/validation split is described, and the only probabilistic intervals shown are for Prophet. This makes the central methodological contribution non-reproducible and the comparison among models unverifiable.
minor comments (4)
  1. [§2.1] The raw filename 'data19 23 raion oblast updated batch2.csv' appears in the text; this is informal and should be replaced with a standardized data descriptor or a DOI/repository link.
  2. [Conclusion] The sentence 'All models agree on the broad trend of stabilized or declining Russian equipment loss rates' is contradicted by the same paragraph's statement that 'XGBoost forecasts ... anticipate a modest uptick in losses during the second half of 2025.' 'All models agree' should be qualified.
  3. [Figure 1 caption] The caption says 'observed values, excluding the partial and potentially anomalous months of June and July 2025', but the ARIMA text elsewhere excludes only June 2025. Clarify which months are excluded and whether the plotted observed series includes July 2025.
  4. [General] Several references are cited in the text but not fully formatted (e.g., 'Boulanin et al., 2020' with a URL in the reference list, and no access date; 'Ward et al., 2013' lacks volume/issue details). Also, citations for model descriptions (e.g., LSTM, TCN) are repeated inconsistently in Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; forecasts extrapolate a self-described partial OSINT series, but no claim reduces by construction to its inputs.

full rationale

The paper's derivation chain is not circular in the sense defined by the review criteria. There are no self-citations, no imported uniqueness theorems, and no fitted parameters relabeled as predictions. The central forecasts are standard time-series extrapolations of the WarSpotting count series, and the paper nowhere defines a derived quantity in terms of the forecast target. Its own 'Limitations of OSINT-Based Loss Data (WarSpotting)' section (pp.16-18) concedes that the series is a visually confirmed lower bound, lagged by weeks, spatially skewed, and vulnerable to platform moderation, and the LSTM section explicitly states that 'current visibility of confirmed losses is primarily shaped by reporting practices rather than combat shocks.' This is a serious construct-validity threat: the forecasts may predict OSINT confirmation processes rather than true battlefield attrition, and the agreement across five models on the same preprocessed series is not independent confirmation. But a measurement-validity problem is not circularity unless the paper defines Russian equipment losses as WarSpotting counts and then presents the forecast as independent derivation of those same counts; it does not. The exclusion of June/July 2025 on data-quality grounds is a subjective preprocessing choice, yet the resulting plateau forecasts are not forced by that exclusion in any equation-level or definitional sense. No specific reduction of a prediction to its input could be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim rests on data-quality assumptions and unstated model hyperparameters, plus a post hoc exclusion of months that is not justified with metrics.

free parameters (7)
  • Excluded months (June and July 2025) = June only in §2.1; June and July in ARIMA section
    Post hoc removal of allegedly incomplete reporting months; changes all forecast levels and is inconsistently described.
  • ARIMA order = (1,1,1)
    Author-chosen, no model selection criterion given.
  • TCN architecture parameters
    Kernel size, dilation, channels, and layers not reported.
  • LSTM architecture parameters
    Layer count, units, sequence length, and dropout not reported.
  • XGBoost hyperparameters
    n_estimators, max_depth, learning rate, and lag features not reported.
  • Prophet settings
    Changepoint prior scale, seasonality modes, and holiday regressors not reported.
  • Train/validation split
    No holdout procedure described; evaluation is visual on the training fit.
assumptions (3)
  • domain assumption WarSpotting counts are a meaningful proxy for actual equipment losses
    Implicit in all sections; contradicted by the paper's own Limitations section (structural undercounting, temporal lag, spatial skew).
  • ad hoc to paper The incomplete reporting months (June-July 2025) should be excluded and the remaining series is representative
    Introduced in §2.1; no data-quality metrics given; inconsistent month ranges.
  • domain assumption Historical loss dynamics will continue through end of 2025
    Standard forecasting stationarity/extrapolation assumption; fragile for conflict data where offensives cause regime shifts; the paper even notes XGBoost contradicts the trend.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forecasting Russian Equipment Losses Using Time Series and Deep Learning Models." pith.science (2026). https://pith.science/paper/27CHHKPP

@misc{pith2026250907813,
  author       = {Pith},
  title        = {Pith review of: Forecasting Russian Equipment Losses Using Time Series and Deep Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27CHHKPP}},
  note         = {Machine review of arXiv:2509.07813}
}
read the original abstract

This study applies a range of forecasting techniques,including ARIMA, Prophet, Long Short Term Memory networks (LSTM), Temporal Convolutional Networks (TCN), and XGBoost, to model and predict Russian equipment losses during the ongoing war in Ukraine. Drawing on daily and monthly open-source intelligence (OSINT) data from WarSpotting, we aim to assess trends in attrition, evaluate model performance, and estimate future loss patterns through the end of 2025. Our findings show that deep learning models, particularly TCN and LSTM, produce stable and consistent forecasts, especially under conditions of high temporal granularity. By comparing different model architectures and input structures, this study highlights the importance of ensemble forecasting in conflict modeling, and the value of publicly available OSINT data in quantifying material degradation over time.

Figures

Figures reproduced from arXiv: 2509.07813 by the authors.

Figure 1
Figure 1. ARIMA Forecast of Monthly Tank Losses (Excluding June–July 2025) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Monthly Predicted Tank Losses Using TCN Forecast (Jul 2025 – Jan 2026) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. TCN Forecast of Daily Tank Losses, June 2025 Excluded (2022–2025) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Monthly Equipment Loss Forecast (All Types) via XGBoost [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: XGBoost Forecast of Daily Equipment Losses (Jul–Dec 2025), June Omitted [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: XGBoost Forecast: Daily Tank Losses Only (Jul–Dec 2025) [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: 30-Day Equipment Loss Forecast using LSTM with Weekday Metadata [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: 7-Day Equipment Loss Forecast (LSTM) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: 70-Day Daily Loss Forecast: LSTM Model with Seasonal Memory [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Prophet Model: Forecast of All Equipment Losses (2022–2026) [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Prophet Model Zoomed on 2024–2025 Loss Forecast [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Agentic and Generative AI for Open-Source Intelligence and Cyber Investigations: Taxonomy, Evaluation, Challenges, and Future Directions

    cs.CR 2026-07 conditional novelty 6.0 of 10

    Across 74 OSINT/CTI AI studies, hallucination is widely named but end-to-end measured in only one non-reproducible system, so a human–AI co-pilot is the most defensible near-term architecture.

Reference graph

Works this paper leans on

11 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Z., & Koltun, V

    Bai, S., Kolter, J. Z., & Koltun, V. (2018). An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.International Conference on Machine Learning, 576–585

    work page 2018

  2. [2]

    Bandara, K., Bergmeir, C., & Smyl, S. (2020). Forecasting daily demand for perishable bakery products using machine learning: A comparison of xgboost, prophet, and lstm. International Journal of Forecasting,36(1), 204–221

    work page 2020

  3. [3]

    Boulanin, V., Carlsson, M., Goussac, N., & Kelley, K. (2020). Artificial intelligence and the future of conflict. https : / / www . sipri . org / publications / 2020 / other - publications / artificial-intelligence-and-future-conflict

    work page 2020

  4. [4]

    Hegre, H., & Sambanis, N. (2006). Sensitivity analysis of empirical results on civil war onset.Journal of Conflict Resolution,50(4), 508–535. https : / / doi . org / 10 . 1177 / 0022002706289303

    work page 2006

  5. [5]

    J., & Athanasopoulos, G

    Hyndman, R. J., & Athanasopoulos, G. (2021).Forecasting: Principles and practice(3rd ed.). OTexts. https://otexts.com/fpp3/

    work page 2021

  6. [6]

    Muchlinski, D. (2021). Using machine learning to predict civil war: Pitfalls and prospects. Political Analysis,29(3), 287–308. https://doi.org/10.1017/pan.2020.24

    work page doi:10.1017/pan.2020.24 2021

  7. [7]

    A., G¨ artner, T., & Meyer, C

    Mutschke, P., Flach, P. A., G¨ artner, T., & Meyer, C. M. (2022). Forecasting conflict with machine learning: The state of the art.Journal of Peace Research,59(4), 504–517. https://doi.org/10.1177/00223433221087529

    work page doi:10.1177/00223433221087529 2022

  8. [8]

    Raleigh, C., Linke, A., Hegre, H., & Karlsen, J. (2010). Introducing acled: An armed conflict location and event dataset.Journal of Peace Research,47(5), 651–660. https://doi. org/10.1177/0022343310378914

    work page doi:10.1177/0022343310378914 2010

Show all 11 references
  1. [9]

    J., & Letham, B

    Taylor, S. J., & Letham, B. (2018). Forecasting at scale.The American Statistician,72(1), 37–45

    2018

  2. [10]

    D., Greenhill, B

    Ward, M. D., Greenhill, B. D., & Bakke, K. M. (2013). Systemic peace forecasting: The future of armed conflict.International Studies Review,15(4), 622–645. 20

    2013

  3. [11]

    Zhang, G. P. (2003). Time series forecasting using a hybrid arima and neural network model. Neurocomputing,50, 159–175. https://doi.org/10.1016/S0925-2312(01)00702-0 21

    2003 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.