Sentio

Falsifiable signal research for Borsa Istanbul (BIST) - is the edge real, or am I fooling myself?

Sentio tests whether a BIST signal survives honest scrutiny instead of just looking good in a backtest. Pre-registration is enforced in code, the pipelines are look-ahead-safe, the data is survivorship-clean, and the validation engine splits the data along whichever axis overfitting likes to hide in.

Across everything I tested, no deployable edge survived. In a market this efficient that's the expected result, and the value here is an honest map of what the data can and can't support, not a wall of green equity curves.

How it works

• Pre-registration is mechanical. Each hypothesis is frozen in a Stage-0 JSON before any result is computed. The engine refuses to run if that file is missing or has drifted, so the goalposts can't move after the fact.
• Null results are the product. Three axes (cross-sectional factors, timing signals, event tilts) are tested and written up whether they work or not. Most don't. Those reports are the real output.
• The validation engine splits where overfit hides. Name-split for cross-sectional factors, purged/embargoed CPCV for timing signals, intra-regime holdout for forward persistence (see below).
• It admits when it can't decide. Every verdict carries a confidence flag, and CONFOUNDED fires automatically when the test window is single-regime. A bare pass with no qualifier is never emitted.

Two tracks

• Anchor (the deployed idea): a low-cost, fully-invested smart-passive strategy that beat the TL-deposit benchmark in real (CPI-deflated) terms over 2019-2026. It lives in the companion repo ballast-bist.
• Lab (this repo): the infrastructure for testing whether any edge exists above that anchor. So far, none does.

The invariant linking both: a signal is never expressed as a cash gate. An idle signal means a fully-invested equal-weight position, not cash. The rule comes from a post-mortem that showed forgone beta, not bad stock selection, was the dominant damage channel.

Data

Daily BIST prices, 2019-2026, 681 symbols, survivorship-clean: delisted tickers stay in, acquired via the official BIST free dataset, so tests don't silently exclude companies that went bankrupt. Seven years covers the modern post-2018 regime; it isn't deep history, and that bounds the statistical power (see Scope & limits).

Validation engine (src/engine/)

A general-purpose, post-hoc-auditable strategy validation tool that enforces research discipline in code rather than convention.

Mode	Question answered	Method
Mod-A (name-split)	Is this cross-sectional overfit to specific tickers?	Disjoint name halves, same window; conjugate agreement across R=50 seed-fixed splits
Mod-B (temporal CPCV)	Is this timing-signal overfit across time?	Embargoed combinatorial purged cross-validation (Bailey & López de Prado)
Mod-C (intra-regime holdout)	Does a factor persist forward within one regime?	Train/holdout split with construction-window embargo

Statistics: Rank IC / IC-IR with Newey-West HAC standard errors; real CSCV-PBO; a Deflated Sharpe bound with honest trial-count binding from Stage-0; market-neutral residuals via rolling-β neutralization; a benchmark floor of real active return vs max(CPI, policy-rate).

Reliability hardening: every conjugate result carries an agreement_confidence qualifier (HIGH / LOW / CONFOUNDED). An iteration lockbox seals a held-out subset with a content hash and enforces single-shot consumption; tampering is visible in git history.

The harness() entry point, the frozen SplitSpec, and the full Section-7 output vector are documented in the operator guide and typed in src/engine/contracts.py.

Research findings

No deployable persistent edge was found across the three tested axes. Cross-sectional factors hit a cost wall (~42 bp realistic round-trip) and regime instability; timing signals failed to clear the significance bar across four independent measurement attempts; event-driven tilts are in forward measurement (recorder live since 2026-06-01, verdict pending). The engine's first real-data exam also surfaced a structural BIST constraint: the survivorship-honest liquid universe (~38 names at the 10M TRY ADV floor over seven years) is too small for the name-split conjugate test at full history. That is a documented property of the market, not a flaw in the method.

Every directive and its verdict is in the research registry; full reports live in docs/research/.

Engineering

Area	Status
Tests	2,088 across 138 files; zero-regression policy, CI blocks merge on any failure
Type checking	mypy strict on src/engine/; older edge modules on a baseline ignore-list being paid down
Linting	ruff clean across src/ and tests/
Look-ahead safety	forward returns computed as return[t+1] on weights[t]; verified by a golden byte-repro gate
Architecture	import-linter contracts block cross-layer imports in CI
CI tiers	architecture → integration → lint → full regression → security; all must pass before merge

Repository structure

sentio/
├── src/
│   ├── engine/        validation engine (harness, contracts, config, Mod-A/B/C)
│   ├── backtest/      backtesting engine + López de Prado statistical validation stack
│   ├── signals/       6-layer signal engine (L1-L6) + regime gate + strategist
│   ├── screening/     factor measurement engines (D-2xx series: event / exposure / trend)
│   ├── analytics/     IC calculator, Brinson attribution, NAV tracker, XBRL scorer
│   ├── data/          fetchers, scrapers, parsers, DataHub router
│   ├── risk/          position sizing, Kelly, drawdown, circuit breaker, stop calculator
│   ├── nlp/           FinBERT / VADER sentiment analyzers
│   └── utils/         logger, weight validator, OS state manager
├── tests/             2,088 tests: unit · integration · architecture invariants · hygiene
├── examples/rry1008/  red-team scripts: graveyard-factor + adversarial overfit probe
├── docs/
│   ├── README.md      documentation guide: what lives where, and why
│   ├── adr/           architecture decision records (immutable once accepted)
│   ├── research/      research reports (RR-xxx, D-xxx, NRR-xxx)
│   ├── yol1/          pre-registrations, results, graveyard
│   ├── specs/         component specifications
│   ├── guides/        operational guides (runbook, architecture, how-tos)
│   ├── disciplines.md research disciplines register (DISC-xx)
│   └── RESEARCH_REGISTRY.md   master index of all research reports
├── scripts/           daily_update.py, health_check.py, build_clean_universe.py
├── config.yaml        runtime config (universe, scanner parameters)
└── pyproject.toml     tool config (mypy, ruff, import-linter contracts)

Getting started

Requirements: Python 3.11+, dependencies in requirements.txt.

git clone https://github.com/caganco/sentio
cd sentio
pip install -r requirements.txt
cp .env.example .env          # keys all optional for offline runs

# Full test suite (runs fully offline)
python -m pytest tests/ -q --tb=short
# Expected: 2,088 passed, 4 skipped (golden gate: needs a local data snapshot)

# Red-team the engine, no external data required
python examples/rry1008/run_part1_known_answer.py

Environment variables (see .env.example):

• EVDS_API_KEY - TCMB EVDS3 macro data (free registration at evds2.tcmb.gov.tr)
• FMP_API_KEY - supplementary market data
• ANTHROPIC_API_KEY - optional; enables the Strategist narrative layer

The engine and test suite run fully offline; API keys only matter for live data pulls. How to attach a signal, write a Stage-0, and read the output vector is in the operator guide.

Scope & limits

Stated up front rather than left to be discovered:

• Trial counting. A Deflated Sharpe is only as honest as its trial count. N is bound per hypothesis-family from Stage-0; the program-level count across 200+ commits is deliberately not collapsed into one number. The research registry is the audit ledger for that.
• One market, one person, no live trading. Solo work on BIST only. Nothing here places orders; there is no execution layer and no live-slippage realization. It's a research instrument by design, not a deployment.
• Shallow history. 2019-2026 is barely more than one regime. The ongoing disinflation gives a genuine prospective out-of-sample window, but today the power is whatever seven years buys.
• Uneven type-checking. The engine core (src/engine/) is mypy-strict; some older edge modules are still on a baseline ignore-list being paid down.

Methodology references

• Harvey, Liu & Zhu (2016). … and the Cross-Section of Expected Returns. Review of Financial Studies 29(1), 5-68. [Multiple testing in factor research; the t > 3 threshold.]
• Bailey & López de Prado (2014). The Deflated Sharpe Ratio. Journal of Portfolio Management. [Deflated Sharpe; the CSCV / PBO framework.]
• McLean & Pontiff (2016). Does Academic Research Destroy Stock Return Predictability? Journal of Finance 71(1), 5-32. [Post-publication decay of anomalies.]

Going deeper

• Operator guide: attach a signal, write a Stage-0, read the output vector
• Research registry: every directive and its verdict
• Architecture and decision log
• ballast-bist: the deployed smart-passive anchor

---

Solo research project. Every finding, including the null ones, is documented in full.

Not affiliated. This is an independent, non-commercial research project with no connection to, sponsorship by, or endorsement from Borsa Istanbul A.S. "Borsa Istanbul" and "BIST" are trademarks of their respective owner, used here only descriptively to identify the market this research studies.