signal-platform/docs/data-integrity-plan.md

Data & signal integrity — implementation plan

Status: closed, 2026-09-07. Steps 1, 2 and 3a ran; Step 3b was not started because 3a did not clear its bar. Findings are in research/survivorship-audit.md, research/tr-rank-agreement.md and research/tr-total-return-ab.md. The follow-on work this raised is in data-reliability.md.

Two questions, answered in order, with the production work made conditional on the answers.

Objective

1. Measure how the backtest changes when removed index constituents are restored. This is a survivorship sensitivity check, not a complete bias estimate.
2. Check whether total-return momentum changes the daily admission or ordering of names. Build the production path only if this screening test finds changes.

Guiding principle: KISS — measure before building. Every earlier draft of this plan started with a schema change (adj_close, then tr_factor), an ingestion maintenance path, a snapshot migration and a benchmark column, in order to deliver a signal change whose effect nobody has measured. Both questions above can be investigated by research scripts against a snapshot copy. Step 1 needs one small, explicit replay option to include entries immediately before a terminal event. No production schema or ingestion changes are needed for either research step. Step 3 is not started until Step 2 finds a ranking or admission difference.

Why the research scripts need none of that machinery: the whole forward-anchored factor design exists because production ingest is incremental, so back-adjusted prices drift between fetches. A research script fetches each ticker's whole history in one request, and within a single response the adjustment is internally consistent. No column, no anchor, no maintenance.

---

Step 1 — How sensitive is the backtest to omitted constituents?

Arm A uses the universe already present in the base snapshot. Arm B adds the names that were removed from the S&P 500 since the snapshot's start — including the ones that failed — and lets the strategy trade them. Use the same configuration, cadence and evaluation dates in both arms. This is deliberately an expanded-universe experiment, without historical membership gating.

Files

• [NEW] scripts/data/sp500_removals.json — flat list, hand-curated from Wikipedia's S&P 500 changes table, checked into git: [{"symbol": "SIVB", "removed_on": "2023-03-15", "reason": "receivership"}, ...] Reasons: receivership | bankruptcy | acquisition | market_cap. No additions, no rename map, no membership intervals — see Not doing.
• [NEW] scripts/prep_survivorship_snapshot.py — makes both arm files from a base snapshot, leaving the base untouched:
  1. Copy base → survivorship-a.sqlite and survivorship-b.sqlite.
  2. Schema sync on both: add every column present in Base.metadata but missing from the file, nullable. The snapshots on disk predate migrations 026 and 032, so select(Ticker) against them fails today with no such column: tickers.cik — and _get_ticker (price_service.py:22) is a full-entity select that query_ohlcv → backtest_service._fetch_columns reaches. Derive the column list from the metadata, not a hardcoded list, so migration 036 doesn't break it again.
  3. Into arm B only: for each removed symbol with no bars, fetch daily bars from Alpaca (it still serves delisted symbols — verified for SIVB, FRC, ATVI, TWTR) and insert the ticker row. Restrict downloaded bars to the base snapshot's date range and use the same price adjustment as its existing bars. Log fetched, already-present and unavailable symbols; an unavailable failure must be called out as missing coverage, not silently treated as having no strategy exposure. removed_on is an index-removal date, not necessarily a delisting date. Nothing is inferred from missing bars: delisted_on / delisted_reason are written only where the dataset carries a confirmed terminal event. A series that merely ends early leaves them NULL and is reported as an unresolved ending status — as built, that is CMA, SEE, HBI and DISH. Series are shortened only at a reviewed cutoff listed in the dataset (SBNY 2023-03-10, INFO 2022-02-25). Elsewhere a run of ≥5 zero-volume sessions is a tripwire, not a rule: the symbol is left out of the arm and reported for review rather than silently truncated.
  4. Terminal mark, arm B only: for a confirmed terminal event, append one synthetic bar on the next SPY trading session after the last real bar, with o=h=l=c=0.01, volume=0. A mark is written only when the retained final bar equals the confirmation's reviewed last_real_bar; a disagreement writes no mark and is reported. Write a sidecar JSON mapping symbol to terminal-bar date for the runner. The portfolio sim already fills a gap through the stop at the open (backtest_service.py ~2280), so the position resolves at ~-100% with no change to its stop-fill logic. Treat 0.01 as a near-total-loss assumption, not an observed execution. For acquisitions, retain the last available close as an approximation; this experiment does not model cash/stock consideration or release of capital on the acquisition date. This matters: failures halt mid-collapse (SIVB's last exchange bar is 106.04 on the way to zero), so without it arm B spares exactly the trades it exists to measure. Appending the bar alone is insufficient: _replay_ticker currently stops at n - HORIZON, where HORIZON = 30. That excludes signals in the last 29 real bars even after the synthetic bar is added. Implement the explicit replay option below before interpreting the audit. Only positions still open at the halt should lose nearly everything; positions stopped or timed out earlier should retain their normal exits. volume=0 is safe — compute_volume_profile guards its only divisor (indicator_service.py:298-300).
• [NEW] docs/research/survivorship-audit.md + an entry in docs/research/README.md.

Small replay change required

• Add optional --terminal-events <json> to run_backtest_snapshot.py. Pass the map explicitly through run_backtest and _replay_and_signals to _replay_ticker, including the multiprocessing path. Default to no events. Do not infer terminal events from low prices or zero volume.
• Validate the whole map before dispatch and fail the run on any mismatch. _replay_ticker rejects a bad date, but both dispatch paths swallow per-ticker exceptions so a long replay survives one unreadable name — which would turn a stale events file into a quietly incomplete report. _validate_terminal_events raises TerminalEventError before any work starts, so no report is written.
• For a mapped ticker, validate that the terminal date matches its final bar. Keep the existing cadence and starting index, but allow signal dates through the last real bar instead of stopping 30 bars early. Never generate a setup on the synthetic bar. Unmapped tickers retain the existing replay bounds.
• Use the existing forward bars and outcome evaluation. For the long-only audit, a setup with a positive stop resolves no later than the terminal bar. Do not pad the history with 30 invented trading bars or turn unresolved shortened histories into normal 30-day outcomes.
• Keep terminal prices out of later signal generation; their role is resolving existing positions. Preserve normal stop, target and time exits before them.

Focused validation

Add regression tests alongside the existing backtest tests:

1. A qualifying long setup five sessions before a terminal event is emitted with the option enabled and exits at 0.01 if still held. Use a controlled setup fixture so the test does not depend on whether SIVB happens to qualify.
2. A position that exits before the terminal event keeps that earlier exit.
3. No setup is emitted on the terminal bar; an ordinary ticker and a run without the option retain their existing replay behavior.

Before the full runs, smoke-test the prepared files with _fetch_columns for one existing and one restored ticker. Confirm matching base-symbol bars between arms, matching configuration and date bounds, and no modification of the base file.

Running it

python scripts/prep_survivorship_snapshot.py --base backtest_snapshots/prod.sqlite
python scripts/run_backtest_snapshot.py backtest_snapshots/survivorship-a.sqlite --out reports/backtest-YYYYMMDD-survivorship-a.json
python scripts/run_backtest_snapshot.py backtest_snapshots/survivorship-b.sqlite --terminal-events backtest_snapshots/survivorship-b-terminal-events.json --out reports/backtest-YYYYMMDD-survivorship-b.json

No new runner: run_backtest_snapshot.py already runs the production sim offline and reports CAGR, max drawdown, Sharpe and total return. Compare the same production portfolio-monitor strategy and all lookback in both reports, not whichever variant performs best. Record B minus A for these metrics, trade counts, evaluated dates, restored-symbol coverage and terminal-event assumptions. A failed ticker replay or missing production report row makes the run incomplete.

What it does and doesn't measure

An expanded-universe sensitivity analysis, not a lower bound. It restores omitted names but does not remove addition lookahead. Still-trading former members remain eligible after removal: run_backtest does not gate on delisted_on, and an index-removal date would not be the right delisting field anyway.

Changing the universe changes percentile ranks and competing admissions. Either correction can raise or lower this strategy's realized returns; do not claim that deletions are the larger source of bias or that the true drag must exceed B minus A. A small delta does not establish that survivorship bias is small. Report it as the sensitivity observed under these assumptions. A complete point-in-time audit remains separate work, not a requirement for finishing this experiment.

Base snapshot

• prod.sqlite: 2021-06-24 → 2026-07-02, 90MB, the ~505 production names.
• research.sqlite: 2016-01-04 → 2026-07-17, 678MB, 4,650 names.

Use prod.sqlite: the highest-signal events (SIVB, FRC, SBNY, ATVI — all 2023) are inside its window, the copies are cheap, and it has no completion manifest to keep in sync. The pre-registration then states a 2021–2026 window, not a ten-year one. Verify the actual file's bounds and record its hash before running. Do not silently substitute research.sqlite: its much broader universe answers a different comparison. Keep the first experiment on the production snapshot.

---

Step 2 — Would total return change any decision?

Use a conservative ranking screen, not a portfolio simulation. Compare the full ordered momentum-admitted list on every daily decision date. Do not call a top-15 list the traded set: production also applies setup qualification, current holdings, available slots and other admission constraints. A difference in this screen is a reason to investigate actual trades in Step 3, not proof of changed P&L.

File

• [NEW] scripts/research_tr_rank_agreement.py — use arm A's universe and dates:
  1. One Adjustment.ALL fetch and one Adjustment.SPLIT fetch per symbol, plus SPY, over the full history. Single fetch each, so each series is internally consistent. Align by date, never zip by index. Cache the responses locally for repeatable runs. Require matching dates in both adjustments; report and resolve missing downloads before applying the gate. Insufficient initial history is expected and should be recorded, with equal eligibility in both arms. Use only bars at or before each decision date.
  2. At each SPY trading date with sufficient history, compute residual 12-1 momentum both ways with the existing momentum_service.compute_residual_12_1_momentum (price-return series against nominal SPY; total-return series against total-return SPY), percentile them with momentum_service._percentiles, and take the admitted set at the production min_momentum_percentile.
  3. Compute compute_realized_vol_6m and its percentiles from the SPLIT series once per date and hold that leg fixed in both arms. This isolates the change to momentum. Order the entire admitted lists by blend_strategy_rank, with symbol as a deterministic tie-breaker. Do not truncate to max_positions: a lower-ranked name may be the first actionable setup when others fail gates or are already held. Iterate symbols in the same sorted order in both arms before calling _percentiles so tied inputs do not create artificial changes.
  4. Save per-date results and an overall/per-year summary: number of evaluated dates, fraction with different admitted sets, fraction with different ordered lists, and mean and maximum set disagreement. Define set disagreement as len(A symmetric_difference B) / len(A union B), or zero for two empty sets. Report Spearman correlation as a diagnostic only. Include the first few changed dates with the changed names to make the result inspectable.

Validation: identical input series must produce no differences. A fixture where the admitted members stay the same but two names swap order must register a changed ordered list. A difference on only one date must still trigger the gate. Write results to docs/research/tr-rank-agreement.md and index it in the research README, including data coverage, parameters and the exact report/cache paths.

Pre-registered decision rule

• No ordered-list differences on any evaluated daily date → stop. Record "no admission or ordering differences observed in this window under the tested configuration." This is historical evidence, not a guarantee about future decisions. No schema or ingestion changes are needed.
• Any ordered-list difference → Step 3. This intentionally conservative rule can flag names that would never actually trade. Step 3's paired portfolio replay determines whether the difference reaches actual admissions and P&L.
• Incomplete downloads or no evaluable dates → inconclusive. Fix coverage before deciding. Never turn missing data into agreement.

There is no median-based tolerance: a zero median can hide differences on nearly half the dates. Checking daily full lists avoids adding a second portfolio engine to this screening script while retaining changes below the nominal top 15.

---

Step 3 — Production total-return ranking (conditional)

Gate cleared 2026-09-07 — Step 3 is now live work. The screen found ordered admitted lists differing on 1,009 of 1,009 evaluated dates, with the shift correlating +0.79 with trailing dividend yield (results). That is a ranking difference, not a P&L difference; Step 3's job is to find out whether it reaches trades.

Step 3a — Paired portfolio replay (do this first, no schema)

Different ranks are not different trades. Before any schema or ingestion work, replay the production book twice and measure actual admissions and P&L. If the books are identical, the basis cannot matter and Step 3b never starts.

[NEW] scripts/run_total_return_ab.py, sourcing everything from artifacts that already exist — backtest_snapshots/survivorship-a.sqlite for bars and backtest_snapshots/tr-rank-cache.sqlite for the two adjusted series.

1. Generate setup candidates once. One replay pass produces the candidate set; both arms reuse it. Only momentum and what depends on it — the momentum percentile, qualification against min_momentum_percentile, and blend_strategy_rank ordering — are recomputed per arm. Setup geometry (entry, stop, target, R:R) is generated once and never differs.
2. Both arms read the same cache. Control takes SPLIT stock series against SPLIT SPY; treatment takes ALL stock against ALL SPY. Neither arm reads a freshly fetched series, and neither mixes cache with snapshot closes for the momentum input — a refreshed-vs-snapshot difference would confound the basis difference being measured.
3. Everything else is pinned. Execution uses the snapshot's nominal bars in both arms; the volatility leg, the non-momentum gates, capacity (max_positions), risk per trade and the exit policy are identical. Each book then carries its own resulting holdings — a divergent admission changes which slots are occupied on later dates, and that path dependence is part of what is being measured.
4. Fixed up front: universe (arm A's 505 symbols with bars), window (2021-06-24 → 2026-07-02), cadence, and the production config as read from the snapshot. Only bars at or before each signal date are used, in both arms.
5. Saved to reports/tr-total-return-ab.json: per-arm trade count, CAGR, max drawdown, Sharpe, total return, win rate; annual returns for both arms and the count of years treatment beats control; and the entry-level diff — dates and symbols entered by one arm and not the other, in both directions.
6. Tests (tests/unit/test_total_return_ab.py): an identical-input parity test — feeding the same series to both arms must produce byte-identical books, so any observed difference is attributable to the basis and not to replay nondeterminism; and a controlled case where a hand-built momentum difference flips one name across the cutoff and the resulting admission actually changes, proving the harness can see what it is looking for.

Reading the result. Different trades establish that the basis is production-relevant. They do not authorize a rollout: that is the non-inferiority decision below, taken on the paired numbers, and Step 3b starts only after the paired results have been reported.

Step 3a result (2026-09-07): the basis reaches trades — ~30% of entries differ. Which basis is better is unresolved: a paired block bootstrap on the date-aligned daily return streams (5,000 replicates, 21- and 63-session blocks, seed fixed beforehand, both reported) puts ΔCAGR at −3.6/−3.8pp with 95% intervals spanning zero, ΔSharpe ≈ −0.11, and ΔMaxDD −1.0/−1.8pp in treatment's favour. Production stays unchanged; 3b not started. Results.

Step 3b — Production storage and ingestion (only if 3a shows real differences)

The design points already settled, worth carrying forward:

• ohlcv_records.tr_factor, Float, nullable, no default and no server_default — a default of 1.0 makes "never backfilled" indistinguishable from "no dividends" and would silently stamp a price-return ranking as total return.
• Forward-anchored: f = 1.0 at max(2016-01-04, ticker's first bar), compounded as f_t = f_{t-1} * (R_t / R_{t-1}) where R = all_close / split_close. The ratio is anchor-invariant only within one API response — never persist R and compare across days.
• upsert_ohlcv must not clobber a stored factor: coalesce(excluded.tr_factor, ohlcv_records.tr_factor). Daily ingest re-fetches the boundary bars, so overlapping writes are routine.
• Daily ingest pulls a ~5-bar overlap under both adjustments to extend the factor; a clean-ratio mismatch on the overlap bars means a split, which needs a full-history refetch (the current incremental ingest never restates history).
• One basis for the whole cross-section, never mixed, stamped in the rank log.
• Beta leg: if the stock series goes total return, SPY must too, in the live path and the backtest, or beta picks up SPY's ~1%/yr yield.
• A/B through the existing matrix on CAGR and max drawdown with per-year win/loss counts. Reuse the same setup candidates, nominal execution bars, core gates, volatility leg, capacity and exit policy; change only momentum and its dependent admission/ranking fields. Replay each book with its own resulting holdings and compare actual entry dates/symbols as well as P&L. Do not substitute top-15 lists for the existing portfolio admission logic. Non-inferiority, not superiority: total return is the definition 12-1 momentum is built on, so an in-sample shortfall shouldn't lock in the basis we believe is measured wrong.

---

Not doing

• Chart total-return toggle. The candlestick chart is an execution chart: stops, S/R levels, entry and paper-trade markers are all nominal. Rescaling the candles draws prices that never traded next to a stop that did.
• [D] dividend markers. Cosmetic, and they need tr_factor to exist. Revisit after Step 3, if Step 3 ever happens.
• Point-in-time addition gating — interval map, reverse-chronological replay, rename map and removal gating are deferred from this sensitivity experiment. Their relative impact is unknown; a small Step 1 delta does not rule them out. Note if a point-in-time audit is ever done: the ranking cross-section must be filtered before _assign_*_percentiles (backtest_service.py:4198-4200), not only at admission in _simulate_portfolio (4257), or the percentile cutoff is still computed against a non-PIT universe.
• Prune protection and the add_ticker regex. Only needed if delisted names are written to the production database. The audit works entirely in snapshot copies. (bootstrap_universe(prune_missing=True) cascading away a still-trading ex-constituent's history is a real latent risk — a separate ticket, not this one.)
• benchmark_prices.tr_close. Only needed for Step 3. Note for then: store a forward factor like the OHLCV side, not a back-adjusted close — refresh_benchmark_prices writes a row only when close changed and is called with variable spans, which would leave different anchors either side of the last refreshed window.
• Migrating the snapshots in place. The prep script syncs its own copies. research.sqlite stays byte-identical, so its completion manifest and the reports that cite it stay valid.