Large Quantitative Models for Treasury: A Comparison

The Expanding Model Landscape

CAPIX’s AI-enhanced cashflow forecasting is built on a hybrid model — LSTM, Transformer and XGBoost working together, feeding an autonomous agent layer that we increasingly run on local hardware. That architecture did not come from picking one algorithm and stopping. It came from evaluating a wide field of Large Quantitative Models (LQMs) — the family of AI and statistical models purpose-built for forecasting and risk on structured, numerical, time-based data — and understanding what each one is, and isn’t, good at.

This article works through the main alternatives we evaluated, and explains why our answer to “which model is best for cash flow forecasting and risk management” is a considered combination rather than a single winner.

What Counts as a Large Quantitative Model?

A Large Quantitative Model, in the treasury sense, is not the same thing as a large language model. It is a model — sometimes small, sometimes genuinely large — trained on numerical, time-series and tabular data to forecast a quantity or estimate a distribution of outcomes: a cash balance, a currency exposure, a probability of a liquidity shortfall. Some of the models below are decades old (ARIMA); some are recent deep learning architectures; a few are pre-trained “foundation” models for time series that only emerged in the last two years. Large language models do have a role in our stack, but as a reasoning layer on top of these forecasts, not as the forecaster itself — more on that below.

The Candidates

Classical statistical models — ARIMA, SARIMA, GARCH

The oldest tools in the box remain relevant. ARIMA/SARIMA provide fast, transparent baselines for trend and seasonality. GARCH models are the standard for volatility clustering — the tendency of FX and rate volatility to arrive in bursts — which makes them a natural fit for risk metrics rather than cash flow point forecasts.

  • Strength: Fully transparent, cheap to run, well understood by auditors and regulators.
  • Limitation: Struggles with multiple currencies, irregular events and nonlinear relationships between drivers.

Gradient boosting — XGBoost, LightGBM

Tree-based ensembles remain hard to beat for structured, tabular data — budget overrides, calendar effects, one-off events — where relationships aren’t strongly sequential. Fast to train, robust to missing data, and explainable via SHAP values.

  • Strength: Quick to deploy, highly interpretable, low compute cost.
  • Limitation: Weaker at capturing long-range temporal dependencies than sequence-native architectures.

Prophet and NeuralProphet

Meta’s Prophet and its neural successor NeuralProphet handle trend, seasonality and holiday effects out of the box with minimal tuning. We tested Prophet against our cashflow data in late 2024 and NeuralProphet through early 2025, and both proved useful as fast, interpretable baselines — particularly where cash flows show strong weekly, monthly or quarterly cycles.

  • Strength: Excellent for seasonal patterns, easy for a treasury analyst to sanity-check.
  • Limitation: Less accurate than deep learning on complex, multi-currency, multi-driver series.

Recurrent deep learning — LSTM, GRU

LSTM and GRU networks are purpose-built for sequential dependencies, which is exactly the shape of accounts payable/receivable history. This is the temporal-forecasting core of our existing hybrid model.

  • Strength: Strong at capturing longer-term temporal patterns across variable time horizons.
  • Limitation: Computationally heavier to train, and needs a reasonable volume of historical data to generalise well.

Attention-based / Transformer models — Temporal Fusion Transformer, Informer, PatchTST

The Temporal Fusion Transformer (TFT) and related architectures apply attention mechanisms to multi-horizon, multivariate forecasting — cash inflows, outflows, FX rates and macro indicators together — with attention weights that offer a degree of interpretability. This is the other half of our current hybrid model, and consistently the architecture that surfaces as the strongest single choice for multi-currency, multi-horizon forecasting.

  • Strength: State-of-the-art accuracy on complex, multivariate, multi-horizon series; scalable.
  • Limitation: Needs more data and tuning than LSTM or gradient boosting; heavier compute footprint.

Probabilistic autoregressive — DeepAR

Amazon’s DeepAR produces full probability distributions rather than single point forecasts, making it well suited to risk-aware forecasting — giving treasury teams a range of plausible outcomes rather than one number.

  • Strength: Native uncertainty quantification, useful directly for risk bands.
  • Limitation: Tied most naturally to a managed SageMaker deployment; less flexible outside that ecosystem.

Pure neural forecasters — N-BEATS, N-HiTS

These architectures (available through libraries such as Darts and Nixtla’s NeuralForecast) forecast purely from the time series itself, without hand-engineered features — fast to train and surprisingly strong on pure trend/seasonality extraction.

  • Strength: Lightweight, fast, no feature engineering required.
  • Limitation: Doesn’t naturally incorporate external drivers like FX rates or budget overrides.

Emerging time-series foundation models — TimeGPT, Chronos, Moirai, Lag-Llama

The newest frontier: large models pre-trained across huge cross-domain time-series corpora, offering zero-shot forecasting on data they’ve never seen, in the same way a large language model generates text without task-specific training. We are evaluating these as a promising complement to our custom-trained models, particularly for fast baseline forecasts on new entities or currencies with limited history — a good candidate for the kind of local, in-house inference we discussed for the NVIDIA DGX Spark.

  • Strength: Forecasts with little or no task-specific training data.
  • Limitation: Still maturing for enterprise treasury use cases; less proven track record than the models above.

Large language models as a reasoning layer — ChatGPT, DeepSeek, Grok-class models

Large language models are not built to output an accurate numeric cash forecast, and we don’t ask them to. Where they do add real value is as the reasoning layer behind our autonomous treasury agents — explaining forecast drivers in plain language, reasoning about scenarios, and deciding what action a policy permits.

  • Strength: Natural-language explanation, scenario reasoning, decision logic.
  • Limitation: Not designed for, and not reliable at, precise numeric time-series prediction.

Comparison at a Glance

Model family Best for Key strength Key limitation
ARIMA / SARIMA / GARCH Volatility & baseline trend Transparent, cheap, audit-friendly Weak on multi-currency, nonlinear drivers
XGBoost / LightGBM Structured/tabular drivers Fast, explainable (SHAP) Limited long-range temporal modelling
Prophet / NeuralProphet Seasonality-heavy series Simple, interpretable Lower accuracy on complex series
LSTM / GRU Sequential cash flow history Captures temporal dependencies Heavier training, needs history depth
Temporal Fusion Transformer Multi-horizon, multi-currency State-of-the-art accuracy Data- and compute-hungry
DeepAR Probabilistic / risk bands Native uncertainty quantification Best suited to managed cloud deployment
N-BEATS / N-HiTS Fast pure time-series forecasts Lightweight, no feature engineering No native external drivers
TimeGPT / Chronos / Moirai Zero-shot / new series Little training data required Still maturing for enterprise use
LLM reasoning layer Explainability & agent decisions Natural-language reasoning Not a numeric forecaster

What About Risk Management Specifically?

Cash flow forecasting and risk management ask different questions of a model. Forecasting asks “what is the most likely cash position?” Risk management asks “how wrong could that be, and what’s the tail scenario?” That distinction matters when choosing a model:

  • Volatility and FX risk are best captured by GARCH-family models layered on top of currency exposure data.
  • Liquidity shortfall risk benefits from probabilistic forecasters (DeepAR, quantile-output TFT) that produce confidence bands rather than a single number.
  • Stress testing and scenario analysis benefit from Monte Carlo simulation over the distributions these models produce, plus an LLM reasoning layer to narrate what a scenario means for treasury policy.

A model chosen purely for point-forecast accuracy will typically under-serve the risk management side of the equation, and vice versa.

So, Which Model Is Best?

There isn’t a single model that wins on forecasting accuracy, interpretability, risk quantification and multi-currency handling all at once — and we’d be sceptical of any vendor claiming otherwise. Across every AI system we consulted during our own research, and across our own testing, the same conclusion keeps recurring: the best result comes from a layered hybrid, not a single model.

CAPIX’s current recommendation, and what underpins our production forecasting engine, is:

  1. Temporal Fusion Transformer + LSTM for the core multi-horizon, multi-currency point forecast.
  2. XGBoost for structured drivers — budget overrides, calendar effects, event flags — feeding into the hybrid as a feature layer.
  3. Prophet / NeuralProphet as a fast, interpretable baseline to sanity-check the deep learning output and catch obvious seasonal misses.
  4. GARCH and quantile/probabilistic forecasting as the dedicated risk layer, producing volatility estimates and confidence bands rather than point forecasts.
  5. An LLM reasoning layer on top of all of the above, explaining forecast drivers and powering the decision logic behind our autonomous agents.

If forced to name a single strongest model for the forecasting core specifically, the Temporal Fusion Transformer is our answer — it consistently handles the multi-horizon, multi-currency shape of treasury cash flow data better than any single alternative. But it is the combination, not the individual model, that makes the difference in production.

Where This Is Headed

We continue to evaluate the emerging foundation models for time series — TimeGPT, Chronos, Moirai — as a faster path to forecasting for new entities and currencies with thin historical data. Running these larger models efficiently is part of why we are bringing AI compute in-house with the NVIDIA DGX Spark: as the model stack grows, so does the value of owning the hardware it runs on.

Learn More

To discuss which combination of models fits your organisation’s cash flow forecasting and risk management needs, contact us for a consultation.

CAPIX – empowering better decisions through intelligent treasury solutions.