Simulacrum’s mission is to build AI agents that see the future, empowering people to navigate uncertainty and make better decisions. We envision a world shaped by foresight—where AI empowers humanity to plan for its future. This note describes the technical problem Simulacrum aims to solve, the solution we propose, and evidence that our solution is effective.

The problem: Existing AI forecasting systems are narrow and fundamentally static.

Existing AI forecasting systems are fundamentally narrow (i.e., limited in scope and unable to generalize across domains) and static (i.e., cannot adapt their forecasts to shifts in the distributions of their predictand). These models’ forecasts perform well within narrow domains, and under the assumption that future data will be distributed as past data.

For instance, in recent years, AI systems such as GraphCast [1] have achieved superior accuracy in predicting weather over a discretized grid of the world—a narrow domain as forecasts apply only to a discrete grid—for a forecast horizon of up to 10 days—a static domain as the distribution of weather data over the next 10 days is likely to behave similarly to the distribution of past data. However, the same models fail to forecast 1) storm and rainfall intensity due to their narrowness (a higher resolution grid is needed for such forecasts), and 2) decadal weather projections due to staticness (atmospheric changes that are not captured in past data heavily impact long-term forecasts) [2]. Weather forecasting systems are only one of the myriad of forecasting systems that behave this way from economics to physics.

The solution: AI Superforecasters

Solving this problem is Simulacrum’s north star, and amounts to development of an AI superforecaster, which at a high-level is a general purpose forecasting system which consistently ranks amongst the top percentiles of all forecasters in terms of forecasting score (e.g., Brier score [3]) across a large and diverse collection of benchmarks (e.g., stock market forecasting) over long time horizons (e.g., decades) [4]. The development of a superforecaster would suggest that the problem before us is solved, as by definition as superforecaster generalizes across domains (i.e., superior performance in diverse benchmarks) and can adapt its forecasts to distribution shifts (i.e., consistent superior performance across long time horizons). By bringing to existence a superforecaster Simulacrum will have achieved its mission, and by doing the work to reach widespread technological adoption we will have achieved our company’s vision.

Our plan to develop a superforecaster will unroll in two technical development phases. First, we will build general-purpose AI forecasting systems, also known as universal forecasters, that can forecast data in any static domain. These systems will perform accurately in short-horizon forecasting problems but their performance will degrade over long-time horizons. Once our universal forecasters have achieved superior performance over existing forecasters, using the infrastructure we have developed, we will develop AI forecasting systems, also known as structural (or hybrid or theory-guided) universal forecasters, which can adapt to distribution shifts and make accurate long-term forecasts. We believe that the structural universal forecaster obtained at the outcome of this second phase will become a superforecaster.

Phase 1: Universal forecasters

A forecaster is a system which outputs future values of a predictand given past-data. Existing forecasters are overwhelmingly narrow, e.g., a price forecaster cannot forecast weather. As a remedy, to the narrowness of forecasters, a nascent literature in deep learning on time-series foundation models has aimed to train general purpose neural forecasters, also known as universal forecaster [5], which take as input any past time-series data (i.e., a context) and forecast the associated future values.

In the last two years, significant progress has been made in training universal (neural) forecasters, but the performance of these models, despite their scale and complexity, remains still on-par with the simplest forecaster in certain forecasting benchmarks [6]. In our paper introducing our time-series foundation model, the Likelihood Aligned Forecast Network (LAFN) [7], as evidence that we can build a highly accurate universal forecaster, we demonstrate that our foundation model achieves superior or on-par performance compared to existing forecasters across a variety of benchmarks, at a fraction of the number of parameters of existing universal forecasters.

Phase 2: Universal Structural Forecasters

Unfortunately, data-driven deep learning approaches are likely to fail at making accurate forecasts in long-term horizons. The reason for this failure is attributed to 1) neural networks’ high sample complexity (i.e., massive data requirement) without enough inductive biases as otherwise they can learn spurious relations [8], and 2) the inability of solely data-driven methods to adapt to distribution shifts (i.e., setting where past data is not representative of future data) [9].

A nascent literature in structural (or hybrid or theory-informed) forecasters [10] has sought to remedy these issue by augmenting the training of neural networks with structural constraints. In Physics, these models have been called physics informed neural networks [11], and have for instance outperformed solely data-driven weather forecasters [12]. In economics, these models have been called structural models [13], and are the state-of-the-art forecasting models used at central banks [14]. In a 2024 ICLR paper written by members of Simulacrum [15], we provide a mathematical and algorithmic framework to build universal structural models, i.e., general purpose structural models, and provide as proof-of-concept that we can successfully complete this phase, experiments on real-world electricity market data where structural model trained with our framework outperforms widely-used forecasters.

Where are we and where are we heading?

Of utmost importance to development of a superforecaster, is the development of a robust forecaster evaluation framework. As a result, today, as we are releasing the first iteration of our universal forecaster LAFN, we are also open-sourcing our internal forecaster evaluation framework Tempus Bench [16]. Our first forecaster is robust, efficient and outperforms existing models on 20+ benchmark forecasting tasks, but it is not perfect. Over the years to come, in line with the roadmap outlined in this paper, we will release increasingly better forecasters and enhance our evaluation framework Tempus Bench to reach our goal of developing a superforecaster. When we achieve this goal, our superforecaster will map our future, and our decisions will be increasingly informed by it, and in many ways it will become an image of the future itself, a simulacrum.

REFERENCES

[1] Lam, Remi, et al. “Learning Skillful Medium-Range Global Weather Forecasting.” Science, vol. 382, no. 6677, 2023, pp. 1416-1421.

[2] Leffer, Lauren. “AI Weather Forecasting Can’t Replace Humans—Yet.” Scientific American, 9 Jan. 2024, www.scientificamerican.com/article/ai-weather-forecasting-cant-replace-humans-yet/.

[3] Brier, Glenn W. “Verification of Forecasts Expressed in Terms of Probability.” Monthly Weather Review, vol. 78, no. 1, 1950, pp. 1-3. American Meteorological Society, doi:10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2.

[4] Tetlock, Philip E., and Dan Gardner. Superforecasting: The Art and Science of Prediction. Random House, 2016.

[5] Woo, Gerald, et al. “Unified Training of Universal Time Series Forecasting Transformers.” arXiv, 2024, arXiv:2401.53140.

[6] Liang, Yuxuan, et al. “Foundation Models for Time Series Analysis: A Tutorial and Survey.” Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024.

[7] Goktas, Denizalp, et al. “Likelihood-Aligned Forecast Networks.” arXiv, 2025.
(Replace ID once finalized, if needed.)

[8] Zhang, Chiyuan, et al. “Understanding Deep Learning Requires Rethinking Generalization.” International Conference on Learning Representations (ICLR), 2017, arXiv:1611.03530.

[9] Koh, Pang Wei, et al. “WILDS: A Benchmark of In-the-Wild Distribution Shifts.” Proceedings of the 38th International Conference on Machine Learning (ICML), 2021, pp. 5637-5664. arXiv:2012.07421.

[10] Karniadakis, George Em, et al. “Physics-Informed Machine Learning.” Nature Reviews Physics, vol. 3, no. 6, 2021, pp. 422-440.

[11] Raissi, Maziar, Paris Perdikaris, and George Em Karniadakis. “Physics-Informed Deep Learning (Part I): Data-Driven Solutions of Nonlinear Partial Differential Equations.” arXiv, 2017, arXiv:1711.10561.

[12] Verma, Yogesh, Markus Heinonen, and Vikas Garg. “CliMODE: Climate and Weather Forecasting with Physics-Informed Neural ODEs.” arXiv, 2024, arXiv:2404.10024.

[13] Keane, Michael P. “Structural vs. Atheoretic Approaches to Econometrics.” Journal of Econometrics, vol. 156, no. 1, 2010, pp. 3-20.

[14] Angrist, Joshua D., and Jörn-Steffen Pischke. “The Credibility Revolution in Empirical Economics: How Better Research Design Is Taking the Con out of Econometrics.” Journal of Economic Perspectives, vol. 24, no. 2, 2010, pp. 3-30.

[15] Goktas, Denizalp, et al. “Efficient Inverse Multiagent Learning.” arXiv, 2025, arXiv:2502.14160.