ALVIN DECISION MODEL HANDOFF

Research Summary

把 DM-001、DM-002、DM-003、MA150 單因子實驗與 QQQ / QLD / TQQQ benchmarks 放在同一份可稽核的研究地圖中。這是 evidence handoff，不是獲利承諾或部署核准。

RESEARCH EVIDENCE DM-001 · FROZEN DM-003 · NOT FROZEN DM-002 · ARCHIVED

下載 Decision Model v1 Matched Results Research Status

EXECUTIVE SUMMARY

目前證據說了什麼

所有比較都保留原始研究狀態；沒有把 research architecture 包裝成 Candidate，也沒有用不同期間的結果做不公平排名。

DM-001 v0.4FROZEN MODEL

QQQ / TQQQ MA200 Hysteresis

三者中 matched comparison 的最高策略報酬，但承受較深 drawdown。它是現行 frozen model；canonical replay 是公平比較層，不覆寫 native frozen contract。

DM-002ARCHIVED / CLOSED

Intraday QQQ → QLD Research

廣泛研究 scale、impulse、ORB、microstructure、cross-asset 與 breadth 後，continuous 2021–2025 經濟性失敗。沒有 v0.8、Candidate 或 registry entry。

DM-003 H0NOT FROZEN

QQQ Trend → QLD / CASH

長期 trend mechanism 有 3/4 independent representations 通過，但 SMA100/150/200 coarse-family gate 判定 isolated；11/12 freeze gates 通過仍不足以 freeze。

DM-001 MA150EXPERIMENT ONLY

One-factor Signal Substitution

aggregate CAGR、DD 與 Calmar 改善，但優勢集中於 2020；排除 2020 後 MA200 略勝，早期與 recovery risk 更差。Classification：mixed / regime-dependent。

DM-001

MA200 hysteresis

Close > 1.04 × MA200 → TQQQ；Close < 0.97 × MA200 → CASH；中間維持 prior regime。

DM-003 H0

SMA150 binary trend

completed canonical QQQ close 高於 SMA150 → next-open 100% QLD；否則 CASH。研究架構，尚未 frozen。

DM-002

Intraday program

QQQ signal、QLD execution、regular session、flat overnight。最終不是單一策略失敗，而是完整 validation chain 未完成。

FULLY MATCHED CANONICAL REVALIDATION

2010-02-11 → 2025-12-31

相同 $100,000 初始資本、canonical next-open execution、2 bps/side、35% primary illustrative tax。Hold benchmarks 採同一 common period。

Metric	DM-001	DM-003	QQQ Hold	QLD Hold	TQQQ Hold
Pre-tax CAGR	32.6369%	24.2665%	19.2574%	32.5875%	42.3003%
35% after-tax CAGR	23.4293%	17.1378%	16.3030%	29.0896%	38.5105%
40% after-tax CAGR	21.9652%	16.0600%	15.7734%	28.4525%	37.8184%
Final wealth @35%	$2,832,773	$1,233,880	$1,101,371	$5,774,930	$17,680,700
Max drawdown	−57.8005%	−41.3000%	−36.6927%	−64.3325%	−81.5773%
Calmar	0.5646	0.5876	0.5248	0.5065	0.5185
Worst year	2022: −35.8467%	2022: −21.1500%	2022: −32.1824%	2022: −60.1256%	2022: −78.7460%
Recovery sessions	128	125	518	628	765
Changes / year	1.38	7.30	0.13	0.13	0.13

Return

DM-001

策略間報酬較高，35% after-tax CAGR 高 DM-003 6.2915pp。

Drawdown

DM-003

比 DM-001 少 16.5005pp 最大回撤；比 QLD Hold 少 23.0325pp。

Passive upside

TQQQ Hold

最高 wealth，同時承受 −81.58% 最大回撤與 765-session recovery。

No universal winner

Different tradeoffs

報酬、回撤、稅負、turnover 與 execution sensitivity 必須一起讀。

Stress-year evidence

2020DM-001 +77.8998%

DM-003 +54.9274%；DM-001 保留更多 upside。

2022DM-003 −21.1500%

DM-001 −35.8467%；DM-003 bear protection 較好。

2025DM-003 +23.0785%

DM-001 +12.2371%；不同 trend horizon 產生差異。

RESEARCH STATUS

證據、限制與下一步邊界

DM-001 v0.4FROZEN MODEL

現行 Alvin handoff installed strategy。Native contract remains authoritative。

DM-002ARCHIVED — CLOSED

2021–2025：pre-tax CAGR +0.42%、after-tax −0.39%、Max DD −31.74%、PF 1.003。

DM-003PROMISING / NOT FROZEN

3/4 mechanism representations pass；coarse SMA robustness gate fails。

MA150MIXED / REGIME-DEPENDENT

值得獨立驗證的 experimental horizon，不取代 MA200。

Four-source canonicalization

Tiingo、Yahoo、Alpaca 與 Norgate 的可用重疊通過 predeclared reconciliation gates。TQQQ level 差異由 split/share-basis presentation 解釋，不是 return disagreement。

Canonical data 在 performance 前 freeze
Norgate licensed overlap 未外推到其 coverage 之外
Raw/licensed data 不在網站或 Git artifact

DM-001 native vs canonical

Canonical agreement 99.2745%，仍有 29 disagreement sessions，transition Jaccard 55.56%。原因是 native Tiingo adjClose 的 backward dividend adjustment 與 canonical capital-only signal contract 不同。

差異已解釋，但未消失。Canonical replay 用於公平比較，不得覆寫 frozen native DM-001。

Key takeaways

DM-001 是較高報酬、較高風險、低 turnover 的 frozen architecture。
DM-003 提供較低 drawdown 與稍高 Calmar，但尚未通過全部 freeze gates。
DM-002 的 intraday evidence 不支持 deployment；archive 是研究結論，不是隱藏失敗。
MA150 的 aggregate 改善不足以證明跨 regime 優越，不能取代 MA200。
Hold benchmarks 顯示 leverage 的 wealth upside 與 drawdown/recovery cost 必須同時呈現。

Authoritative artifact references

research/dm001_vs_dm003_fully_matched_canonical_v2/ research/DM002_FINAL_ARCHIVE.md research/dm003_canonical_h0_mechanism_robustness_v2/ research/dm001_ma150_signal_substitution_v1/

DELIVERY SCOPE

這次交付的是什麼

PRODUCT

Decision Model v1

一個 strategy-agnostic decision engine。它接收經核准的 point-in-time market snapshot，執行已安裝策略，最後產生 portfolio intent。

INSTALLED STRATEGY

DM-001 v0.4

QQQ / TQQQ Regime Model。它是目前安裝在 Decision Model 裡的第一個 Candidate strategy，不是 external product identity。

Decision Model ≠ DM-001 交付的是完整 Decision Model runtime；DM-001 只是 v1 目前唯一安裝的策略。

只有通過 gates 才能加入

DM-002 已封存，DM-003 尚未 frozen；兩者都沒有加入 registry。未來策略仍須完成各自 validation / audit。

正常情況下 Alvin 不需重做

DecisionRequest / PortfolioTargets integration
Risk Gateway
Order Planner
Broker integration

RESPONSIBILITY BOUNDARY

Architecture

Decision Model 負責形成投資組合意圖；Alvin Platform 保留交易授權、風控與執行。

Alvin Platform

↓

Point-in-Time Market Snapshot

↓

DecisionRequest

DECISION MODEL BOUNDARY

Decision Model v1

↓

Snapshot Validation

↓

Strategy Registry

↓

DM-001 v0.4

↓

Allocator

↓

PortfolioTargets / NO_DECISION

ALVIN PLATFORM BOUNDARY

Signal Gateway

↓

Risk Gateway

↓

Order Planner

↓

LEAN / Brokerage

DECISION MODEL 負責

接收 approved point-in-time market data
驗證 snapshot 並執行策略
產生 portfolio intent
回傳 PortfolioTargets 或 NO_DECISION

DECISION MODEL 不負責

Broker credentials 或 Brokerage connectivity
BUY / SELL order submission
Risk limits 與 Order planning
Partial fills、Reconciliation、Kill switch

EXTENSIBILITY

Multi-Strategy Design

Decision Model

Strategy Registry

DM-001 v0.4ACTIVE / INSTALLED

DM-002ARCHIVED / NOT INSTALLED

DM-003RESEARCH / NOT INSTALLED

Replaceable Allocator

External adapter 不含 DM-001 strategy logic，也不直接 import DM001StrategyModule。策略集合被封裝在 Decision Model composition boundary 後方。

1新增 StrategyModule

2註冊 Strategy

3完成 Tests / Audit

4更新 Allocator configuration

5適當地升級 Decision Model version

新增策略不應只是因為策略數量改變，就迫使 Alvin 修改 API、Risk Gateway、Order Planner 或 Broker integration。

CURRENT INSTALLED STRATEGY

DM-001 v0.4 — QQQ/TQQQ Regime Model

Candidate Baseline Frozen v0.4 Independent Audit: PASS WITH WARNINGS

DM-001 是 deterministic、price-based、low-frequency regime strategy。它使用 QQQ adjusted close 與 MA200 判斷市場 regime，再把 portfolio intent 映射為 TQQQ exposure 或 CASH。

Daily EODNo newsNo fundamentalsNo alternative dataNo AI sentimentNo online learningNo InternetNo GPU

MA200 排除當日

MA200(t) = mean[Close(t−200) … Close(t−1)]

只使用 t−1 以及更早的 200 個 QQQ adjusted closes，不把當日 Close(t) 放進比較基準。

BULL

Close(t) > 1.04 × MA200(t)

Regime → BULL；portfolio intent 為 TQQQ target 100%。

BEAR

Close(t) < 0.97 × MA200(t)

Regime → BEAR；TQQQ target 0%，其餘為 CASH。

Hysteresis

0.97 × MA200 ≤ Close ≤ 1.04 × MA200

介於兩個門檻之間，包括 boundary equality，維持 prior regime，避免在均線附近頻繁切換。

Regime	TQQQ target	CASH	Contract meaning
BULL	100%	0%	Valid PortfolioTargets
BEAR	0%	100%	Valid explicit-zero PortfolioTargets
UNINITIALIZED	0%	100%	Valid CASH decision

Valid CASH ≠ NO_DECISION

CASH 是策略做出的有效配置。NO_DECISION 則代表資料完整性、history 或 validation 不足，模型無法安全產生決策。

STATE & DETERMINISM

State Replay

UNINITIALIZED→Replay approved history→Reconstruct prior regime→Current decision

DM-001 的 hysteresis 讓 prior regime 有意義，但 Decision Model v1 不依賴 hidden mutable state。只要提供完整、approved historical snapshot，就能從 UNINITIALIZED deterministic replay 出相同狀態。

same approved snapshot + same model version + same configuration → same decision

Fail closed：不猜測 regime

Insufficient historyTruncated historyDuplicate datesNon-monotonic datesInvalid pricesStale / invalid snapshotData integrity failure

EVIDENCE

DM-001 Validation

32 / 32Original builder / regression

14 / 14Independent Audit

23 / 23Trade reconciliation

16 / 16Real DM-001 internal integration

27 / 27Original synthetic integration

PASSFull frozen QQQ observation parity

PASSDeterministic replay

ONE-FACTOR RESEARCH COMPARISON

DM-001 MA150 Experiment

EXPERIMENTAL / NOT FROZEN

只把 frozen DM-001 v0.4 的 MA200 horizon 替換為 SMA150；1.04 / 0.97 hysteresis、TQQQ/CASH、next-open execution 與其他規則全部不變。這不是 DM-001 v0.5，也未取代原始 MA200。

MetricMA200 FrozenMA150 Experiment

Pre-tax CAGR32.6369%34.6904%

35% after-tax CAGR23.4293%25.4251%

40% after-tax CAGR21.9652%23.9431%

Max Drawdown−57.8005%−52.2530%

Calmar0.56460.6639

Worst year2022: −35.8467%2011: −41.0209%

Changes/year1.381.51

Average holding190.3 sessions173.8 sessions

Next-close CAGR impact−0.3462pp−0.7021pp

研究結論：aggregate risk/return 改善，但優勢主要由 2020 貢獻；排除 2020 後 MA150 年化略低於 MA200，且 2011 明顯較差。Classification：MA150 MIXED / REGIME DEPENDENT。原始 DM-001 v0.4 維持 frozen / unchanged。

Frozen snapshot SHA-256 ef86ffd4f646fa7f953f4b3aa2741589aadcb0b256b4062450255a88fe405ad2

Independent Audit: PASS WITH WARNINGS

MIN-001：frozen historical data 沒有保存逐列 QQQ/QQQQ ticker alias 或永久 instrument identity metadata。日期與價格 continuity、transition observations 與所有獨立重算訊號均相符；此 warning 影響 provenance / auditability，並非 signal 或 execution mismatch。

DM-001 仍是 Candidate，不是 Approved strategy。Integration implementation 已與 audited/frozen v0.4 logic 完成 parity validation；frozen baseline、spec 與 audit artifacts 未因 integration 修改。

PACKAGE GATES

Package Validation

Linux AMD64 build/runtimePASS

ARM64 buildPASS

ARM64 emulated runtimePASS

Native ARM64NOT YET VERIFIED

Clean-room packaged testsPASS

Redistribution-safe fixturesPASS

Deterministic replayPASS

External contract strategy-agnosticPASS

Strategy Registry extensiblePASS

Allocator replaceablePASS

DM-001 external couplingNONE

Broker connectivity / CredentialsNONE

Outbound inference networkNOT REQUIRED

ARM64 結果是在 Docker Desktop emulation 下取得；不等於 native ARM64 acceptance。

RESOURCE PROFILE

Performance

LOCAL TECHNICAL MEASUREMENT — NOT PRODUCTION SLA

最新 package evidence：Linux AMD64 container、CPython 3.11.9。Cold time 包含 Docker Desktop container 建立與移除。

Warm internal mean0.886 ms

Warm p992.642 ms

Warm max11.692 ms

Parse + decide + serialize mean1.331 ms

Cold container / CLI mean532 ms

Process max RSS20,320 KiB

Python heap peak430,932 B

GPU requirementNone

START HERE

Alvin Integration Guide

1

下載 package
Decision-Model-Handoff-v1.zip
2

驗證 SHA-256
b39301bfa0606c79e631503e13b9147487a9e5d0c23fae4a15470ad1757c13d1
3

解壓縮並先閱讀
ALVIN_HANDOFF.md
4

Build validation image

docker buildx build --platform linux/amd64 --target validation --provenance=false --load -t decision-model-handoff:v1-validation .
5

Fixture verification

docker run --rm --read-only --network none --cap-drop ALL --security-opt no-new-privileges --entrypoint python decision-model-handoff:v1-validation -B /app/verify_handoff.py
6

Packaged tests

docker run --rm --read-only --network none --cap-drop ALL --security-opt no-new-privileges --entrypoint python decision-model-handoff:v1-validation -B -m unittest discover -s /app/tests -v
7

Run example request

docker run --rm --read-only --network none --cap-drop ALL --security-opt no-new-privileges --entrypoint python decision-model-handoff:v1-validation -B -m decision_model.handoff_cli /app/fixtures/handoff/requests/valid-target.json --snapshot-root /app/fixtures/handoff
8

確認 contract flow
DecisionRequest → Decision Model → PortfolioTargets / NO_DECISION
9

Review provisional assumptions
閱讀 contract/CONTRACT_ASSUMPTIONS.md，把需修改的 platform-side assumptions 回覆給 Alan。
10

Contract agreement 後
Freeze External Contract v1 → next integration package → acceptance → Shadow。每一步皆需另行核准。

INSPECTED ZIP CONTENTS

Package Contents

ALVIN_HANDOFF.mdAlvin 的主要操作與驗收入口

DockerfileProduction / validation multi-stage image

model-manifest.yamlProduct、contract、strategy、runtime 與 gates 的 machine-readable manifest

MODEL_CARD.md用途、限制、determinism、governance 與 data requirements

README.mdPackage 快速說明

contract/DecisionRequest、PortfolioTargets、NO_DECISION v0 schemas 與 assumptions

src/decision_model/Decision Model source、DM-001 module、Registry、Allocator、Snapshot validation、Internal decision 與 External adapter

fixtures/Redistribution-safe synthetic requests、snapshots 與 expected responses

tests/Architecture、runtime、integration、determinism 與 contract tests

verify_handoff.py單一 fixture verification entrypoint

FILE_INVENTORY.txtPackage 內 73 個檔案的 inventory

SHA256SUMS.txtRelevant package artifacts 的 SHA-256 checksums

以上項目均由實際 ZIP inventory 確認；package 不包含 Tiingo raw data、broker credentials 或 Personal Radar research。

EXTERNAL CONTRACT

Contract Status

這一版可以

Docker / package review
Fixture verification
Contract review
Adapter integration exploration

這一版不代表

External Contract accepted
Shadow / Paper / Canary / Live authorized
Deployment authorized
DM-001 promoted to Approved

REQUIRES ALVIN REVIEW

Contract Assumptions Still Open

Request / response fields and optionality

Request ID / Decision ID ownership

Duplicate / idempotency behavior

Snapshot transport與 authentication

Freshness / provenance / failure behavior

target_percent representation與 precision

patch semantics與 explicit zero exit

Expiry / timeout policy

NO_DECISION reason vocabulary

Target architecture / resources

Image registry / signing / SBOM

Digest與 acceptance process

目前 convention 只是讓 package 可以執行與驗證；網站不宣稱 Alvin 已接受任何 provisional assumption。

INTEGRATION REVIEW PACKAGE

下載 Decision Model v1

取得可執行 package、schemas、fixtures、tests 與完整 handoff guide。

Download Decision Model v1

File: Decision-Model-Handoff-v1.zip
SHA-256: b39301bfa0606c79e631503e13b9147487a9e5d0c23fae4a15470ad1757c13d1
Contract: v0-provisional
Purpose: Integration Review

Research Summary

目前證據說了什麼

QQQ / TQQQ MA200 Hysteresis

Intraday QQQ → QLD Research

QQQ Trend → QLD / CASH

One-factor Signal Substitution

MA200 hysteresis

SMA150 binary trend

Intraday program

2010-02-11 → 2025-12-31

Return

Drawdown

Passive upside

No universal winner

Stress-year evidence

證據、限制與下一步邊界

Four-source canonicalization

DM-001 native vs canonical

Key takeaways

Authoritative artifact references

這次交付的是什麼

Decision Model v1

DM-001 v0.4

只有通過 gates 才能加入

正常情況下 Alvin 不需重做

Architecture

Multi-Strategy Design

DM-001 v0.4 — QQQ/TQQQ Regime Model

MA200 排除當日

BULL

BEAR

Hysteresis

State Replay

Fail closed：不猜測 regime

DM-001 Validation

DM-001 MA150 Experiment

Independent Audit: PASS WITH WARNINGS

Package Validation

Performance

Alvin Integration Guide

下載 package

驗證 SHA-256

解壓縮並先閱讀

Build validation image

Fixture verification

Packaged tests

Run example request

確認 contract flow

Review provisional assumptions

Contract agreement 後

Package Contents

Contract Status

這一版可以

這一版不代表

Contract Assumptions Still Open

下載 Decision Model v1