TLDR: A study by Jonathan Teagan utilizes various forecasting models, including deep learning (LSTM, TCN) and statistical methods (ARIMA, Prophet, XGBoost), to predict Russian equipment losses in Ukraine based on open-source intelligence from WarSpotting. The research finds that deep learning models provide stable and consistent forecasts, while XGBoost anticipates a modest increase in losses during the second half of 2025. The study emphasizes the importance of ensemble forecasting and the value of OSINT data, while also detailing its inherent limitations such as undercounting and reporting lags.
A recent study delves into the complex task of forecasting Russian equipment losses during the ongoing conflict in Ukraine, utilizing a diverse array of advanced analytical techniques. Authored by Jonathan Teagan, the research, titled Forecasting Russian Equipment Losses Using Time Series and Deep Learning Models, provides critical insights for military strategists and policymakers by assessing attrition trends and estimating future loss patterns through the end of 2025.
The study leverages daily and monthly open-source intelligence (OSINT) data, primarily from WarSpotting, which meticulously documents visually verified Russian equipment losses. This unprecedented level of detail allows for a granular analysis of material degradation over time, covering everything from tanks to infantry fighting vehicles and support vehicles.
Methodology and Models
To achieve its forecasting goals, the research employs a range of models, from traditional statistical methods to cutting-edge deep learning architectures. These include:
-
ARIMA (AutoRegressive Integrated Moving Average): A classical statistical model, serving as a baseline for its ability to handle linear trends and autocorrelations in time series data.
-
Prophet: Developed by Facebook, this model is known for its robustness in handling irregular data, missing values, and automatically detecting trend changes and seasonal patterns.
-
LSTM (Long Short-Term Memory): A type of deep learning neural network particularly adept at learning long-term dependencies in sequential data, making it flexible for capturing nonlinear behaviors.
-
TCN (Temporal Convolutional Networks): Another deep learning approach that uses dilated causal convolutions to efficiently capture both short-term and long-term dependencies, often providing smoother and more stable forecasts than LSTMs.
-
XGBoost: A powerful gradient-boosted decision tree algorithm, which, while not intrinsically a time series model, excels when time series data is transformed into a supervised learning problem using lagged features.
The dataset used spans from February 2022 to early July 2025. A crucial preprocessing step involved excluding June 2025 due to incomplete reporting, ensuring that the models were trained on reliable and consistent data.
Key Findings Across Models
The study’s findings offer varied perspectives on the future trajectory of Russian equipment losses:
-
ARIMA Forecasts: This model, trained on monthly tank loss data, projects a continuation of stabilized tank losses, anticipating approximately 80–100 vehicles per month through the end of 2025. It provides a conservative baseline, suggesting a plateau in armored vehicle losses.
-
TCN Forecasts: The Temporal Convolutional Network predicts stable tank losses around 100 units per month from July to December 2025. This suggests a saturation effect, where combat intensity and loss rates converge towards a plateau, smoothing out historical volatility.
-
XGBoost Forecasts: In contrast to the stabilization predicted by ARIMA and TCN, XGBoost projects an increase in overall equipment losses from July to October 2025. This could be linked to seasonal offensives or replenishment cycles, with tank-specific losses estimated at 3–5 vehicles per day.
-
LSTM Forecasts: LSTM models, which account for weekday effects in OSINT confirmation (e.g., weekend lulls), tend to produce smoothed predictions. They suggest stable low-loss regimes, implying that current confirmed losses are more influenced by reporting practices than sudden combat shocks.
-
Prophet Forecasts: This model forecasts a slow, gradual decline in daily confirmed losses through 2025. It identifies a logarithmic decay trend, with high losses in early 2022 stabilizing in 2023–2024, consistent with the depletion of armored platforms and adaptations in Russian logistics.
Overall, there is a broad consensus among the models on stabilized or declining Russian equipment loss rates through late 2025, though XGBoost offers a nuanced view with a potential uptick.
Limitations of OSINT Data
The study also critically acknowledges the inherent limitations of OSINT-based data from platforms like WarSpotting. These include:
-
Undercounting: Only visually confirmed losses are included, meaning many losses without public visual documentation go uncounted.
-
Reporting Lag: The date of visual confirmation often lags behind the actual loss event, introducing temporal distortions.
-
Spatial Skew: Data is more prevalent from regions with high drone surveillance or active OSINT contributors, leading to biases.
-
Misclassification: Low-resolution imagery can lead to incorrect identification of equipment types.
-
Denominator Problem: The data tracks losses but not total deployed inventories, making it difficult to assess loss rates proportionally to total forces.
-
Platform Volatility: Changes in social media moderation or censorship can affect data availability.
Despite these limitations, the research underscores the immense value of OSINT data in providing transparency and granularity to conflict analysis, especially when combined with rigorous modeling techniques.
Also Read:
- Understanding AI’s Decisions in Complex Logistics Systems
- Advanced Burned Area Mapping Using AlphaEarth and Siamese Networks
Conclusion and Future Directions
The study concludes that deep learning models, particularly TCN and LSTM, offer the most temporally consistent forecasts at the daily level, while Prophet excels in identifying smoother strategic trends. ARIMA serves as a reliable statistical baseline. The variety of insights from different models highlights the importance of ensemble forecasting—combining multiple perspectives—when dealing with the inherent uncertainties of conflict modeling.
Future research could enhance these models by integrating geospatial predictors, unit-level loss affiliations, and event-based triggers (like the start dates of major offensives) to improve both tactical interpretability and operational relevance.
