Models that reason, inside the mandate.
The intelligence layer turns market structure, filings and language into proposals, and shows its work. Models propose; the kernel, the mandate compiled, decides what is permitted; every signal carries its reasoning to the ledger.
Intelligence the mandate governs.
Models propose inside limits you authored. The mandate compiles to policy the kernel enforces before any order exists, so a model can never widen its own envelope.
- Generate and rank signalsAUTO
- Retrain a model in shadowAUTO
- Flag low-confidence language outputNOTIFY
- Promote a challenger modelAPPROVAL-GATED
- Change the mandate a model runs underAPPROVAL-GATED
Mandate Studio · Illustrative sample data
Three model families, designed to work as one.
No single method captures every market. Gradient-boosted and deep models for cross-sectional structure, language models for filings and transcripts, statistical time-series models for regime and volatility: their outputs are combined and weighted by the conditions each handles best, not averaged blindly.
- 机器学习模型梯度提升树以及深度与循环网络，设计用以学习横截面与序列结构，并配以持续的特征工程与再训练。
- 语言模型LLM 管线设计用以解读申报文件、财报会议纪要、管理层评述与新闻——将非结构化语言转化为结构化、经推理的输入。
- 时间序列模型状态空间、机制切换与波动率模型（ARIMA/GARCH 族系、Kalman 滤波），设计用以追踪结构性变化并预测风险。
- 集成综合贝叶斯模型平均与加权投票，设计用以整合预测、揭示模型间的分歧，并校准为一个带置信度评分的单一信号。
每一个信号都是一条有据可查的管线的终点。
A signal is not a single model call. It passes through ingestion, feature construction and ensemble synthesis, then reaches the kernel as a proposal. Only what the mandate permits can become an order, and the reasoning travels with it.
- 摄取与特征市场、基本面、另类与宏观数据共同输入一个由技术、微观结构、情绪与跨资产信号构成的特征层。
- 集成与综合模型预测经过聚合，相互冲突的信号在不同时间跨度间得以协调，信号强度则依据风险回报评分加以校准。
- The kernel decidesEach candidate is checked against position, concentration, liquidity and policy limits before it can become an order. Allowed or blocked, the outcome is written to the decision ledger with the clause that decided it.
- 排序交付通过筛选的信号在交付时，设计为按序排列，并附带置信度与风险指标、建议头寸规模，以及入场与离场参数。
每一项决策皆设计为可被解释。
Institutional trust requires knowing why. Each signal is designed to carry attribution, plain-language reasoning, a confidence profile and, for any language-model output, a validation record, all written to the ledger and answerable in the Observatory. Opacity is treated as a defect.
- 特征归因SHAP 与 LIME 风格的归因，设计用以揭示哪些输入驱动了每一项预测，并将正向与负向贡献分项列出。
- 白话推理每一个信号皆设计用以携带一段白话理由：关键驱动因素、支持与反对的证据，以及其背后的历史背景。
- 置信度与不确定性模型一致性评分与不确定性量化，设计为随信号一同传递，而不仅是一个点估计。
- LLM 推理防火墙语言模型的输出，设计用以对照结构化数据校验幻觉与逻辑一致性，来源经过追踪，并在低置信度时触发人工复核。
模型在投产之前、之中与之后皆受治理。
Models are designed to clear validation, run in shadow against the incumbent, and stay under monitoring once promoted. The explanation that helps a portfolio manager is the same record an examiner reads.
- 部署前已验证前向滚动与净化交叉验证、泄漏管控、过拟合校验与独立复核，皆在任何内容投产之前便已设计纳入模型开发流程。
- Champion and challengerA new model runs in shadow beside the incumbent. Promotion waits for a named approver to sign it, and the decision is recorded.
- Monitored for driftModel behaviour, assumption drift and degradation are designed to be tracked continuously, with recalibration and rollback triggers.
- Prompt & version controlLLM prompts, model versions, and data are versioned so results are reconstructable.
- A record an examiner can readA model inventory, documented limitations and every explanation are kept on the decision ledger, written once and never edited.
See the intelligence layer in the full system.
It does not run alone: it is one part of an operating system Adaptic runs live on its own capital. Evaluate it on your own mandate in a six-week shadow pilot, with no capital at risk.
Product imagery shows illustrative sample data, not a track record. Adaptic holds no licence in any jurisdiction yet; its proprietary fund trades only its own capital, and it will not manage outside capital until it is authorized. Capability descriptions state what the system is designed to do.