How we model

Zero-shot forecasting · rolling-origin grading · conformal intervals

Use it to forecast demand, prices, load or cash flow, with a range on each number. Pretrained transformers have done for forecasting what large language models did for text: a single model, trained on billions of time points, forecasts a series it has never seen, zero-shot, with no per-series training. We work across the current frontier (Google’s TimesFM¹, a decoder-only model that uses input patching; Amazon’s Chronos², which tokenises series values into a language-model vocabulary; Salesforce’s Moirai; and Nixtla’s TimeGPT) and grade every forecast by rolling-origin cross-validation against a horizon-matched random walk³, publishing the grade with the forecast, the misses included. In vnpredict, fourteen series that revert to their mean blend a local AR(1) model with Chronos-2, two use Chronos-2 alone, and USD/VND uses a robust-drift estimator (a Huber M-estimator) that beats both Chronos-2 and the random walk and carries a conformally calibrated band; on 15 October 2026, 14 of the 17 forecasts beat the random walk, and the other 3 were graded moderate.

In productionvnpredict ↗

Gradient-boosted quantiles · prediction intervals · tabular foundation models

Use it to value or price an asset from its attributes, with a range on each estimate. On structured, tabular data, gradient-boosted decision trees (XGBoost, LightGBM, CatBoost) have been the workhorse for a decade and remain hard to beat. We fit three quantiles¹ (p10, p50 and p90), one model each, so every estimate arrives with an interval, and we measure how often outcomes land inside it. The frontier is tabular foundation models, which perform in-context inference in a single forward pass and now beat tuned trees on small-to-medium data². In vnappraise, three CatBoost models trained at α = 0.1, 0.5 and 0.9 give the band, which covers 77.2% of out-of-sample asking prices against a nominal 80%; TabICL, a tabular foundation model, is available as the alternative.

Research previewvnappraise ↗

Dense + sparse · merged ranking · reranking

Use it to find the passage in your documents that answers a question. Dense embeddings and sparse keyword search fail in opposite directions. A bi-encoder captures meaning (matching “termination of agreement” to “contract cancellation”) and can miss exact identifiers like a statute number or a ticker; BM25¹ nails the exact token and misses the paraphrase. We run both in parallel and merge the ranked lists, for example with Reciprocal Rank Fusion², which combines on rank position and so sidesteps the score incompatibility that breaks naively weighted pipelines. A reranker then scores the top candidates jointly against the query³. In vnadvise, each question is split into sub-queries that run in parallel on semantic and keyword search; the results are merged and reranked, and each claim in the answer cites the passage it came from.

In productionvnadvise ↗

Structured output · tool use · ReAct · web-grounded

Use it to automate document intake, extraction and analysis, with each step checked. We treat LLMs as components inside deterministic, schema-checked pipelines. Output is structured through native tool-use / function-calling: the model returns typed tool_use blocks validated against a schema, so downstream code never parses free text. Agents that need to act use the ReAct¹ loop (interleaving reasoning, tool calls, and observations), grounded in real tools and live web results², with guardrails and validation at every hop. The result is automation you can audit.

LLM steps inall four products ↗

CAPM/WACC · robust statistics · Bayesian · Monte Carlo · conformal

Use it where a figure must follow a published formula that a reviewer can recompute. Modern models sit on top of the quantitative-finance backbone. We use CAPM and WACC for cost of capital and discounting; robust statistics¹ (Huber M-estimators), so a handful of outliers cannot dominate a fit; quantile regression² for distribution-aware estimates; and Bayesian inference and Monte Carlo simulation to carry uncertainty through to the result. Tying it together is conformal prediction³: a distribution-free way to turn any model’s output into an interval with a finite-sample coverage guarantee. In vnvalue, CAPM sets the cost of equity, with beta from VNINDEX regressions and a 2.75% country risk premium, and every figure is a published formula on public filings; in vnpredict, a Huber M-estimator drives the USD/VND forecast inside a conformally calibrated band.

In productionvnvalue ↗