tac-exp-dev/book/README.md

# TradeAC Quant Trading Guide — Table of Contents & Status

A practitioner's guide to quantitative trading written the only way it is worth reading: grounded in a real research loop and a real execution trail. Every number in this book was either reproduced from a recorded TradeAC experiment (MLflow run + traced git branch) **on the clean lake (exp 21+)** or a reconciled post-reset live round, or it is explicitly labeled a hypothesis. See `AGENTS.md` (repo root) for the truth contract; `EVIDENCE.md` for the ledger; `CLAIMS.md` for the proven-vs-hypothesis matrix.

## Evidence boundary and living-document status

- **The clean-lake boundary (2026-08-18, exp 21) is the evidence watermark.** Anything before it — exp 8–18 and their backtests, pre-reset live rounds — is historical context and idea material only, never cited as fact (they were demonstrably inflated by lake data-quality problems, `EVIDENCE#010 → exp 21`). Pre-reset experiments and all opencode chat transcripts (see `data/chat_mining/` and `references/chat-ideas.md`) feed the book's hypothesis pipeline.
- **Every section is living.** As new experimental results land on the clean lake, chapters are updated; a chapter marked `done` is done for its window, not forever.

## What this book is for

A quant-desk reader should be able to act on this book: replicate a signal pipeline, size a book, gate it with risk limits, execute it, and reconcile what actually happened. The book's spine is **how performance improved with research-proved truth** — the actual arc of TradeAC's campaign from a baseline that barely cleared costs to a live, reconciled round.

## How to read evidence tags

- `PROVEN` — reproduced from a recorded run or reconciled round. Citation is an `experiment_id`/`run_id` or `round_id`.
- `HYPOTHESIS` — plausible but not yet reproduced; never stated as fact.
- `REFERENCED` — industry/academic practice; citation is an external source.
- `TODO(evidence-needed: …)` — an open question the desk should settle.

## Table of contents

| # | Chapter | Status | Core experiments cited | Core lesson |
|---|---------|--------|------------------------|-------------|
| 00 | Why a real execution trail matters | drafting | round 3 | A book claims nothing it cannot reconcile |
| 01 | Metrics: the vocabulary of a price series | drafting | exp 21–31 + dataset studies | Every claim reduces to a falsifiable statistic |
| 02 | The research loop: lake → experiment → live | drafting | exp 8–31 | Traceability is the methodology |
| 03 | Baseline and the cost reality | drafting | exp 22–26, 28–31 | A signal that dies after 5bp/15bp is not a signal |
| 04 | Prune, don't add: feature-family ablation | drafting | exp 9, 10, 11, 25, 43 | On a 50-name panel, generic beats model-specific; standalone reversal never existed |
| 05 | Ensembles and the seed-count effect | drafting | exp 12, 28, 34 | Averaging raises ICIR; more seeds raise breadth but not net — cost is the ceiling |
| 06 | The clean-lake reset: data quality as first-order risk | drafting | exp 21–24 | If it doesn't reproduce on clean data, it was noise |
| 07 | Isolation runs: single-variable discipline | drafting | exp 26, 29–31, 33–37, 43 | Most additions fail; the discipline is the value |
| 08 | Portfolio construction: dropout vs the rest | drafting | exp 13, 14, 15, 35, 38, 39, 41 | Turnover-sensitive construction bleeds the edge; weekly recompute wins |
| 09 | The cost/turnover frontier | drafting | exp 26, 39, 41 | Cut turnover before adding signal; weekly rebalance is the proven lever |
| 10 | Risk limits and gates that work | drafting | exp 18, 20, 40, 42 | Limits are a safety net, not alpha; gates churn without signal |
| 11 | Walk-forward re-validation and guard candidates | drafting | exp 52–56 | The edge is a 2025–2026 regime artifact; all 5 guards refuted |
| 12 | Live execution and reconciliation | drafting | exp 27, round 3 | 4.54 bps slippage realized; funnel 10→10→10→9 |
| 13 | Synthesis: how proved truth compounds | drafting | all, exp 33–43, 52–56 | Cost relief > signal; the Q-campaign scoreboard |

Status legend: `drafting` → `in-review` → `done`.

## Per-chapter claim inventory (expected truth status)

Each chapter opens with its claims. The inventory below is the working contract: what the chapter asserts, and what evidence tier it must land in. It is updated as chapters pass their HITL review gate.

### 00 — Why a real execution trail matters
| Claim | Expected status |
|-------|-----------------|
| A book's claims must be reconcilable to a real trail (targets→decisions→fills) | `PROVEN` — round 3 funnel |
| Backtest claims without live reconciliation are hypotheses about execution | `HYPOTHESIS` → settled by round 3 |
| The funnel (targets→decided→placed→filled) is the minimal honesty structure | `REFERENCED` (industry ops practice) + `PROVEN` via tac-rd-book schema |

### 01 — Metrics: the vocabulary of a price series
| Claim | Expected status |
|-------|-----------------|
| Every chapter claim reduces to a statistic computable on the lake (drift, jump, vol, regime, reversion, memory, risk, error, probability, timeline, decay) | `PROVEN` (chapters 03–13) + `HYPOTHESIS` (dataset-study magnitudes, chat-derived) |
| Generic scale-free statistics beat model-specific machinery on a small daily panel | `PROVEN` (exp 23/24/25/29/31) + `HYPOTHESIS` (generality) |
| A statistic is only as good as the falsification it survives (null z-scores, reproduction) | `PROVEN` (exp 21 detection playbook) + `REFERENCED` |
| The strongest single-feature signal (OU z-score) can be worthless inside a rank model — the "OU paradox" | `PROVEN` (exp 25) + open mechanism `TODO(evidence-needed)` |

### 02 — The research loop
| Claim | Expected status |
|-------|-----------------|
| Experiments must be traced: branch + MLflow run + notes (hypothesis before run) | `PROVEN` — traceability loop used on exp 8–31 |
| Pre-registration protects against post-hoc cherry-picking | `REFERENCED` (research practice; see CLAIMS for multiple-testing note) |
| The lake is the single source of bar/feature truth | `PROVEN` — exp 21 showed dirty-lake risk |
| One variable changes per run (isolation); verdicts attributable | `PROVEN` — exp 26→28/29/30/31 design |

### 03 — Baseline and the cost reality
| Claim | Expected status |
|-------|-----------------|
| Baseline 1-day LGB signal is weak on 2026 OOS (RankIC ≈ 0.04, below the 0.2 ICIR noise threshold) | `PROVEN` — exp 8 |
| Costs erase most of the raw edge: +6.2% ann gross → +1.6% net | `PROVEN` — exp 8 |
| A viable signal must clear realistic execution costs | `PROVEN` (exp 8, exp 26) + `REFERENCED` |

### 04 — Prune, don't add
| Claim | Expected status |
|-------|-----------------|
| Dropping model-specific feature families (ou, hmm) improves the rank signal (RankIC 0.030→0.064) | `PROVEN` — exp 9 |
| Adding moment/volatility families regresses the signal (exp 11), same failure mode as ou/hmm | `PROVEN` — exp 11 |
| Adding OU mean-reversion (sp_ou_zscore) hurts on clean data | `PROVEN` — exp 25 |
| Standalone 5d reversal (single feature sp_trend_slope_5) is not learnable — model trains positive IC | `PROVEN` — exp 43 (Q11) |
| More features ≠ better signal on a small cross-section | `HYPOTHESIS` (supported by 3+ runs, still panel-specific) |

### 05 — Ensembles
| Claim | Expected status |
|-------|-----------------|
| 5-seed RankIC ensemble raises net-of-cost performance vs single model on the ablated set | `PROVEN` — exp 12 (pre-clean-lake), re-validated exp 22–24 |
| Seed count is load-bearing: 2 seeds lose to 5 seeds on clean data | `PROVEN` — exp 28 |
| 10 seeds raise rank breadth (RankIC 0.0671, L/S Sharpe 4.58) but the book stays negative net | `PROVEN` — exp 34 (Q02) |
| Ensemble averaging's benefit is separable from feature expansion | `PROVEN` — exp 12 isolation design |

### 06 — Clean-lake reset
| Claim | Expected status |
|-------|-----------------|
| The reference signal did not reproduce on a rebuilt lake (IC 0.035→0.002) | `PROVEN` — exp 21 |
| Data-quality problems had inflated earlier results; post-reset signal is the only valid one | `PROVEN` — exp 21 + exp 22–24 reproduction |
| Signal work must be re-validated after any data rebuild | `PROVEN` (exp 21) + `HYPOTHESIS` for generality |

### 07 — Isolation runs
| Claim | Expected status |
|-------|-----------------|
| Single-variable changes isolate what moved performance | `PROVEN` — exp 26→29/30/31 + exp 33–43 Q-runs design |
| Multi-horizon momentum degrades the reference (net IR 0.21→-1.12) | `PROVEN` — exp 29 |
| Risk-adjusted 22d Sharpe drift (M2): reproduced on the compact set by Q01 | `PROVEN` — exp 30 + exp 33 (Q01) |
| GARCH(1,1) vol-regime features add no signal | `PROVEN` — exp 31 |
| Longer labels raise IC monotonically but net worsens under daily turnover (10d/22d) | `PROVEN` — exp 36/37 (Q04/Q05) |
| Standalone reversal feature does not reproduce | `PROVEN` — exp 43 (Q11) |

### 08 — Portfolio construction
| Claim | Expected status |
|-------|-----------------|
| TopkDropout beats stochastic-control OptimalStopControl on the ensemble signal | `PROVEN` — exp 13, 14 |
| Stop-control constructions churn and bleed costs (cost drag ≈ −11.3pp) | `PROVEN` — exp 13 |
| Weekly rebalance recompute of the daily signal is the campaign's best construction (net +12.51%, IR 1.24) | `PROVEN` — exp 39 (Q07) |
| Fractional-Kelly sizing (exp 15) is refuted on the clean lake (net +1.04%, IR 0.11) | `PROVEN` — exp 38 (Q06) |
| Widening the book (topk 20) adds no edge; long-short top/bottom is destroyed by turnover | `PROVEN` — exp 35/41 (Q03/Q09) |

### 09 — Cost/turnover frontier
| Claim | Expected status |
|-------|-----------------|
| n_drop 2→1 flips net excess from −3.21% to +2.13% with identical signal metrics | `PROVEN` — exp 26 |
| Cost drag is the binding constraint, not signal quality | `PROVEN` — exp 26 (IC/RankIC identical between n_drop variants) |
| Weekly recompute cuts cost drag to ~1.1pp and unlocks +12.51% net | `PROVEN` — exp 39 (Q07) |
| Long-short daily turnover costs 9.7% of NAV ($96.7k); fill rate 0.40 | `PROVEN` — exp 41 (Q09) |

### 10 — Risk limits
| Claim | Expected status |
|-------|-----------------|
| $5M liquidity floor improves net IR 0.81→0.98 and cuts drawdown 7.9%→5.4% | `PROVEN` — exp 18 (pre-clean-lake; see note in chapter) |
| Size/concentration caps hurt by cutting deployed capital | `PROVEN` — exp 18; re-confirmed clean-lake exp 40 (Q08) |
| Entry/risk gates are no-ops when the signal is the bottleneck | `PROVEN` — exp 20 (R2/R3 byte-identical) |
| Exp-18 numbers are not comparable to post-reset runs due to env non-determinism | `PROVEN` — exp 20 R0 note |
| Post-reset A/B: the floor binds but adds no IR edge; DD relief is pure defunding | `PROVEN` — exp 40 (Q08) |
| HMM regime gate meets only the drawdown leg and churns | `PROVEN` — exp 42 (Q10) |

### 11 — Walk-forward re-validation and guard candidates
| Claim | Expected status |
|-------|-----------------|
| The headline results (weekly +12.51%, m2-sharpe22 +6.5%) are 2026-window-specific; walk-forward re-training across 2024/2025 is negative or flat | `PROVEN` — exp 52/53 |
| A and C share identical predictions; the strategy layer alone decides the outcome | `PROVEN` — exp 52 |
| No pre-deployment measurable gate (feature-PSI, label-regime PSI, streaming IC, window length, staleness) selects a profitable year | `PROVEN` — exp 52–56, all 5 guards refuted |
| The edge is a 2025–2026 regime artifact; live capital must be cut until the regime returns | `PROVEN` (walk-forward) + `HYPOTHESIS` (forward-looking) |

### 12 — Live execution and reconciliation
| Claim | Expected status |
|-------|-----------------|
| Live funnel held: 10 targets → 10 decided → 10 placed → 9 filled, 1 cancelled, 1 skipped | `PROVEN` — round 3 |
| Realized slippage ≈ 4.54 bps, estimated cost ≈ $45, turnover 0.74 | `PROVEN` — round 3 metrics |
| Live beats backtest: execution claims trace to round_id, not to backtest | `PROVEN` — methodology |

### 13 — Synthesis
| Claim | Expected status |
|-------|-----------------|
| The largest performance deltas came from data quality, cost/turnover relief, feature pruning, and risk limits — not from adding features | `PROVEN` — composite of exp 9, 18, 21, 26, 39 |
| The campaign's refuted runs (exp 11, 13, 14, 20, 25, 29, 31, Q02–Q06, Q09–Q11) were as valuable as wins | `REFERENCED` + `PROVEN` (they stopped wrong directions) |
| Turnover reduction is the dominant net-performance lever (weekly rebalance +12.51% vs daily −3.21%–+2.13%) | `PROVEN` — exp 26 vs 39 |
| The campaign's headline edges were a 2025–2026 regime artifact, not robust OOS | `PROVEN` — exp 52–56 |
| Generalizability of the 50-ETF panel results is an open question | `HYPOTHESIS` — TODO(evidence-needed: out-of-panel universe) |

## Repository layout

```
book/
  README.md          # this file
  EVIDENCE.md        # ledger: id → claim → source → verified?
  CLAIMS.md          # proven-vs-hypothesis matrix, updated every chapter
  chapters/00-intro.md ...   # one file per chapter
  data/              # ad-hoc validation scripts + outputs
  data/chat_mining/  # raw opencode chat transcripts (idea sources)
  references/chat-ideas.md   # distilled ideas/hypotheses from chats + pre-reset experiments
  references/        # external citations
```

## Open questions for the desk

- `TODO(evidence-needed: a second live round beyond round 3, to confirm slippage and funnel hold under a different market regime)`
- `TODO(evidence-needed: reconcile realized cost against the 5bp/15bp/$5 backtest model over a full position window)`
- `TODO(evidence-needed: whether sp_sharpe_22 still helps when combined with the weekly-rebalance construction of ch. 08)`
- `TODO(evidence-needed: automated lake-integrity check wired into every experiment run, not only on demand)`
- `TODO(evidence-needed: live round under weekly-rebalance construction with risk-limit spec, to confirm safety-net behavior at higher deployed capital)`
- `TODO(evidence-needed: a live window that matches the 2026 label regime, to test whether the edge returns when the regime returns)`
- `TODO(evidence-needed: a causal (no-lookahead) regime-change detector that selects the 2026 window before the fact — none of the five guards did)`

### Settled open questions (no longer active)

- ~~`weekly-rebalance result (exp 39) reproduced on a second window before promotion to a live round`~~ — **ANSWERED (negatively):** Q13 (exp 45) tested weekly on 2025 OOS: net −4.21% IR −0.52. The edge is window-dependent, not robust. `EVIDENCE#037`.
- ~~`long-horizon label (10d/22d) paired with a low-turnover construction`~~ — **ANSWERED:** Q12 (exp 44): 22d+weekly net −4.88% IR −0.566. Q21 (exp 51): 10d+weekly net +1.19% IR 0.148. Both below IR 0.5 acceptance. Weekly is a universal cost lever (~10pp improvement) but the5d label remains the sweet spot. `EVIDENCE#036/042`.
- ~~`out-of-universe (non-ETF) validation of the compact stochastic feature set`~~ — **ANSWERED (negatively):** Q14 (exp 50): RankIC −0.02, ICIR −0.07 on 30 liquid single-stock names. Signal is noise outside the 50-ETF panel. `EVIDENCE#033`.
- ~~`exp 18 risk-limit spec reconciliation — post-reset A/B (exp 40) shows it is a safety net, not alpha`~~ — **ANSWERED:** Q08 (exp 40): $5M floor binds but adds no IR edge (candidate 1.512 < baseline 1.580). DD relief is pure defunding. `EVIDENCE#029`.
- ~~`do the headline results survive walk-forward re-training?`~~ — **ANSWERED (negatively):** exp 52/53/54 re-ran weekly, moments, ndrop2, and m2-sharpe22 across 2024–2026 (plus 2021/2023 label-regime matches). Only 2026 is profitable; all prior years negative or flat. Edge = 2025–2026 regime artifact. `EVIDENCE#043–045`.
- ~~`is there a pre-deployment guard that isolates the profitable regime?`~~ — **ANSWERED (negatively):** feature-PSI, label-regime PSI, streaming IC (`ic_min_rankic`), adaptive short-window, and staleness guards all refuted. `EVIDENCE#043–047`.