trader-cloud-backtest

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Cloud backtest / train (neural-trader on a Managed Agent)

云端回测/训练（基于Managed Agent的neural-trader）

Dispatch a heavy

neural-trader

job to an Anthropic Claude Managed Agent (cloud container) instead of running it locally. See project ADR-117 (recipe + cost rules) and ADR-115 (the

managed_agent_*

runtime).

将大型

neural-trader

任务分发至Anthropic Claude Managed Agent（云端容器）运行，而非本地执行。详情请参考项目ADR-117（方案及成本规则）和ADR-115（

managed_agent_*

运行时）。

When to use this vs

trader-backtest

(local)

何时使用本方案 vs

trader-backtest

（本地版）

Job	Runtime
Quick sanity check; one short backtest (< ~1 min)	local — use the `trader-backtest` skill
Multi-year walk-forward, big Monte-Carlo count, parameter sweep over a grid, or model training (LSTM/Transformer/N-BEATS)	cloud — this skill

Prereq:

ANTHROPIC_API_KEY

(or

CLAUDE_API_KEY

) + Managed Agents beta access. If

managed_agent_*

returns "needs ANTHROPIC_API_KEY", fall back to the local

trader-backtest

skill.

任务类型	运行环境
快速sanity check；单次短时间回测（<约1分钟）	本地 — 使用 `trader-backtest` 技能
多年滚动回测、大规模蒙特卡洛模拟次数、基于网格的参数寻优，或模型训练（LSTM/Transformer/N-BEATS）	云端 — 本技能

前置条件：

ANTHROPIC_API_KEY

（或

CLAUDE_API_KEY

）+ Managed Agents测试版访问权限。若

managed_agent_*

返回"needs ANTHROPIC_API_KEY"，则fallback到本地

trader-backtest

技能。

Steps

步骤

Estimate first. From the job size, print an estimated cost (≈ container-minutes × rate + tokens) — a long sweep is a deliberate choice, not a default.

Provision (or reuse) the container — install neural-trader at container start so the agent doesn't reinstall mid-run:

managed_agent_create({
  name: "nt-cloud",
  model: "claude-haiku-4-5-20251001",            // orchestration only — the compute is the Rust engine, not the LM (ADR-026)
  system: "You operate the `neural-trader` CLI in this container. Run exactly the commands asked, report the metrics, write requested artifacts, then stop.",
  networking: "unrestricted",                     // or "restricted" pinned to your data host
  packages: { npm: ["neural-trader"] },           // add apt:["build-essential"] ONLY if there's no prebuilt NAPI binary for the arch (neural-trader ships prebuilds → usually omit)
  initScript: "npm install -g --ignore-scripts neural-trader >/dev/null 2>&1 || npx -y neural-trader --version >/dev/null 2>&1 || true"
})
→ { sessionId, agentId, environmentId }

For a sweep: create the environment once, run all configs in one

managed_agent_prompt

(one container), not N sessions.

Pre-flight cheap. Before a 1000-path / multi-year run, do a tiny smoke first (1 MC path, ~3 months) — catches a bad strategy name / symbol in seconds:

managed_agent_prompt({ sessionId, message: "Run `npx neural-trader --backtest --strategy <name> --symbol <TICKER> --period <last 3 months> --mc-paths 1`. Just confirm it ran and report the Sharpe. Then stop.", maxWaitMs: 60000 })

If that fails, fix the args before the real run (and

managed_agent_terminate

Run the real job:

managed_agent_prompt({
  sessionId,
  message: "Run `npx neural-trader --backtest --strategy <name> --symbol <TICKER> --period <range> --walk-forward --mc-paths <N>` (for training: `npx neural-trader --train --model <lstm|transformer|nbeats> --symbol <TICKER> --period <range>`; for a sweep: loop the configs and run each). Report: total return, annualized return, Sharpe, Sortino, max drawdown, win rate, profit factor, # trades, 95% CVaR. Write the equity curve to /tmp/equity.csv and the trade log to /tmp/trades.csv. Then stop.",
  maxWaitMs: <generous — minutes>
})
→ { finished, status, stopReason, assistantText (the metrics), toolUses }

finished:false

, follow up with

managed_agent_events({ sessionId })

until idle.

Pull artifacts (if needed):

managed_agent_prompt({ sessionId, message: "cat /tmp/equity.csv" })

managed_agent_events

and read the tool_result.

Ingest locally:

memory_store({ key: "backtest-<strategy>-<ts>", value: <metrics+params JSON>, namespace: "trading-backtests" })

If Sharpe > 1.5:

agentdb_pattern-store({ pattern: "profitable-<strategy-type>", data: "<params + results>" })

Record the run's container time + token cost to the
```
cost-tracking
```
namespace (per ADR-117 — cloud sessions bill until terminated).

Terminate immediately — results in hand:

managed_agent_terminate({ sessionId, environmentId })   → { sessionDeleted: true, environmentDeleted: true }

Never leave an idle billing container. (

ruflo doctor

/ GC catches orphans — #1931.)

先估算成本。根据任务规模，打印预估成本（≈容器分钟数 × 费率 + 令牌消耗）——长周期参数寻优是主动选择，而非默认选项。

创建（或复用）容器——在容器启动时安装neural-trader，避免代理在运行中途重新安装：

managed_agent_create({
  name: "nt-cloud",
  model: "claude-haiku-4-5-20251001",            // 仅用于编排——计算由Rust引擎执行，而非大语言模型（ADR-026）
  system: "You operate the `neural-trader` CLI in this container. Run exactly the commands asked, report the metrics, write requested artifacts, then stop.",
  networking: "unrestricted",                     // 或设置为"restricted"并固定到你的数据主机
  packages: { npm: ["neural-trader"] },           // 仅当对应架构没有预构建NAPI二进制文件时才添加apt:["build-essential"]（neural-trader已提供预构建版本→通常无需添加）
  initScript: "npm install -g --ignore-scripts neural-trader >/dev/null 2>&1 || npx -y neural-trader --version >/dev/null 2>&1 || true"
})
→ { sessionId, agentId, environmentId }

对于参数寻优：只需创建一次环境，在同一个

managed_agent_prompt

（同一个容器）中运行所有配置，而非N个会话。

低成本预校验。在运行1000路径/多年任务前，先执行一次小型冒烟测试（1次蒙特卡洛路径，约3个月）——可在数秒内发现错误的策略名称/交易代码：
```
managed_agent_prompt({ sessionId, message: "Run `npx neural-trader --backtest --strategy <name> --symbol <TICKER> --period <last 3 months> --mc-paths 1`. Just confirm it ran and report the Sharpe. Then stop.", maxWaitMs: 60000 })
```
如果测试失败，在正式运行前修复参数（并执行
```
managed_agent_terminate
```
终止容器）。

运行正式任务：

managed_agent_prompt({
  sessionId,
  message: "Run `npx neural-trader --backtest --strategy <name> --symbol <TICKER> --period <range> --walk-forward --mc-paths <N>`（训练任务使用：`npx neural-trader --train --model <lstm|transformer|nbeats> --symbol <TICKER> --period <range>`；参数寻优任务使用：循环遍历所有配置并逐一运行）。需报告：总收益、年化收益、夏普比率（Sharpe）、索提诺比率（Sortino）、最大回撤、胜率、盈利因子、交易次数、95%条件风险价值（CVaR）。将权益曲线写入/tmp/equity.csv，交易日志写入/tmp/trades.csv。完成后停止。",
  maxWaitMs: <充足时长 — 以分钟为单位>
})
→ { finished, status, stopReason, assistantText (指标结果), toolUses }

若

finished:false

，后续执行

managed_agent_events({ sessionId })

直到容器处于空闲状态。

拉取产物（如需）：执行

managed_agent_prompt({ sessionId, message: "cat /tmp/equity.csv" })

或通过

managed_agent_events

读取tool_result。

本地存储：

memory_store({ key: "backtest-<strategy>-<ts>", value: <指标+参数JSON>, namespace: "trading-backtests" })

若夏普比率>1.5：

agentdb_pattern-store({ pattern: "profitable-<strategy-type>", data: "<params + results>" })

将本次运行的容器时长+令牌消耗记录到
```
cost-tracking
```
命名空间（遵循ADR-117——云端会话在终止前持续计费）。

立即终止容器——获取结果后：
```
managed_agent_terminate({ sessionId, environmentId })   → { sessionDeleted: true, environmentDeleted: true }
```
切勿让计费容器处于空闲状态。（
```
ruflo doctor
```
/垃圾回收机制会清理孤立容器——#1931。）

Cost rules (don't skip)

成本规则（请勿跳过）

Install once (
```
initScript
```
), reuse the environment, batch sweeps into one prompt, pre-flight cheap, terminate eagerly, use Haiku/Sonnet for the agent loop, estimate before kicking off. (ADR-117 §"Cost optimization".)
A cloud backtest that runs for an hour costs an hour of container time + the agent-loop tokens. Be deliberate.

仅安装一次（通过
```
initScript
```
），复用环境，将参数寻优任务批量放入同一个prompt，执行低成本预校验，及时终止容器，使用Haiku/Sonnet作为代理循环模型，启动前先估算成本。（ADR-117 §"成本优化"章节）
运行1小时的云端回测将产生1小时的容器时长费用+代理循环令牌消耗。请谨慎操作。

Quick example

快速示例

managed_agent_create  { "name":"nt-cloud", "model":"claude-haiku-4-5-20251001", "packages":{"npm":["neural-trader"]}, "initScript":"npm install -g --ignore-scripts neural-trader >/dev/null 2>&1 || true" }
  → { sessionId:"sesn_…", environmentId:"env_…" }
managed_agent_prompt   { "sessionId":"sesn_…", "message":"Run `npx neural-trader --backtest --strategy multi-indicator --symbol SPY --period 2020-2024 --walk-forward --mc-paths 1000`. Report Sharpe/Sortino/max-DD/win-rate/CVaR; write /tmp/equity.csv. Then stop.", "maxWaitMs":600000 }
  → { finished:true, status:"idle", assistantText:"<metrics>", toolUses:[{bash:"npx neural-trader --backtest …"}] }

managed_agent_create  { "name":"nt-cloud", "model":"claude-haiku-4-5-20251001", "packages":{"npm":["neural-trader"]}, "initScript":"npm install -g --ignore-scripts neural-trader >/dev/null 2>&1 || true" }
  → { sessionId:"sesn_…", environmentId:"env_…" }
managed_agent_prompt   { "sessionId":"sesn_…", "message":"Run `npx neural-trader --backtest --strategy multi-indicator --symbol SPY --period 2020-2024 --walk-forward --mc-paths 1000`. Report Sharpe/Sortino/max-DD/win-rate/CVaR; write /tmp/equity.csv. Then stop.", "maxWaitMs":600000 }
  → { finished:true, status:"idle", assistantText:"<metrics>", toolUses:[{bash:"npx neural-trader --backtest …"}] }

… memory_store the metrics, agentdb_pattern-store if Sharpe>1.5, record cost …

… 将指标存入memory_store，若夏普比率>1.5则存入agentdb_pattern-store，记录成本 …

managed_agent_terminate { "sessionId":"sesn_…", "environmentId":"env_…" }

undefined

managed_agent_terminate { "sessionId":"sesn_…", "environmentId":"env_…" }

undefined

trader-cloud-backtest

Original

Translation

Cloud backtest / train (neural-trader on a Managed Agent)

云端回测/训练（基于Managed Agent的neural-trader）

When to use this vs
`trader-backtest`
(local)

何时使用本方案 vs
`trader-backtest`
（本地版）

Steps

步骤

Cost rules (don't skip)

成本规则（请勿跳过）

Quick example

快速示例

… memory_store the metrics, agentdb_pattern-store if Sharpe>1.5, record cost …

… 将指标存入memory_store，若夏普比率>1.5则存入agentdb_pattern-store，记录成本 …

trader-cloud-backtest

Original

Translation

Cloud backtest / train (neural-trader on a Managed Agent)

云端回测/训练（基于Managed Agent的neural-trader）

When to use this vs trader-backtest (local)

何时使用本方案 vs trader-backtest（本地版）

Steps

步骤

Cost rules (don't skip)

成本规则（请勿跳过）

Quick example

快速示例

… memory_store the metrics, agentdb_pattern-store if Sharpe>1.5, record cost …

… 将指标存入memory_store，若夏普比率>1.5则存入agentdb_pattern-store，记录成本 …

When to use this vs
`trader-backtest`
(local)

何时使用本方案 vs
`trader-backtest`
（本地版）