FinSight

▶ Live demo: finsight.dexdevs.com — run it in your browser, free offline backend. Browse all 10 portfolio demos via the all demos link.

Turn financial news sentiment into a tradable factor — and prove it isn't luck.

FinSight builds a cross-sectional, dollar-neutral long/short signal from news sentiment, backtests it with realistic transaction costs, and then subjects the result to the statistics that separate a real edge from backtest overfitting: the Deflated Sharpe Ratio, the Probabilistic Sharpe Ratio, and purged + embargoed walk-forward cross-validation (Bailey & López de Prado). A classic news event study cross-checks the signal from an independent angle.

It is offline-first: a deterministic synthetic market embeds a known sentiment→return relationship, so the whole pipeline has ground truth and CI runs green with no API keys and no model downloads. Real backends (FinBERT sentiment, yfinance prices) are opt-in via pip extras.

Most backtests answer "what was the Sharpe?" FinSight answers the question that actually matters: "given how many configurations I tried, how likely is this Sharpe to survive out of sample?"

Demo

$ finsight backtest
FinSight strategy [lexical sentiment / synthetic data]
  IC (rank)      : +0.052
  Ann. return    : +13.60%
  Ann. vol       : 7.05%
  Sharpe         : 1.93
  Sortino        : 2.03
  Max drawdown   : -8.21%
  Avg turnover   : 0.305/day

$ finsight score "Shares plunge as the firm warns and cuts guidance"
[lexical] -1.000 (negative)  'Shares plunge as the firm warns and cuts guidance'

Architecture

flowchart LR
    subgraph Data["Data backend (env-selectable)"]
        SYN[Synthetic market<br/>known signal + AR(1) news]
        CSV[CSV / yfinance]
    end
    subgraph Sentiment["Sentiment backend"]
        LEX[Lexical<br/>finance lexicon · offline]
        FB[FinBERT<br/>optional extra]
    end
    SYN --> NEWS[Headlines]
    CSV --> NEWS
    NEWS --> LEX --> SCORE[Daily asset sentiment]
    NEWS --> FB --> SCORE
    SCORE --> SIG[Cross-sectional z-score<br/>→ dollar-neutral weights]
    SIG --> BT[Backtest engine<br/>shift+1 timing · txn costs]
    BT --> STATS[Stats]
    SIG -.-> ES[Event study · CAR]
    Data -.-> ES
    subgraph STATS["Overfitting-aware statistics"]
        SR[Sharpe / Sortino / drawdown]
        PSR[Probabilistic Sharpe]
        DSR[Deflated Sharpe<br/>N-trial correction]
        CV[Purged walk-forward CV]
    end

Loading

The metric that matters

The strategy is selected from a search over 7 smoothing windows. A naive report would quote the winner's Sharpe and stop. FinSight instead deflates it:

	Real signal	Scrambled (null) signal
Annualised Sharpe	1.93	−1.42
Deflated Sharpe Ratio	0.997	0.000

When the alpha is shuffled away, the same search still finds a "best" configuration — but the Deflated Sharpe correctly collapses to zero. That gap is the whole product.

Full offline eval

Reproduce with python -m evals.harness (writes evals/RESULTS.md):

Metric	Value	Reads as
Sentiment recovery corr	0.84	lexical model vs. latent truth
Information coefficient (rank IC)	0.052	signal vs. next-day return
Annualised return (net)	+13.6%	after 10 bps/turn costs
Sharpe (annualised)	1.93	in-sample, selected config
Walk-forward OOS Sharpe	2.10	purged + embargoed
Probabilistic Sharpe (PSR)	1.000	P(true Sharpe > 0)
Deflated Sharpe (DSR)	0.997	search-adjusted
Event-study long-short CAR (t)	+0.013 (t = 8.45)	independent confirmation

Quickstart

python -m venv .venv && . .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

pytest -q                    # 18 tests
python -m evals.harness      # full eval table -> evals/RESULTS.md
python -m evals.gate         # CI quality gate (floors on every metric)

finsight backtest            # run the strategy
finsight score "..."         # score a headline
finsight eventstudy          # news event study (CAR)

One-command Docker run (offline):

docker build -t finsight . && docker run --rm finsight

Backend matrix

Component	Offline default	Optional real backend	Env var
Sentiment	lexical (finance lexicon)	finbert (pip install "finsight[finbert]")	FINSIGHT_SENTIMENT_BACKEND
Market data	synthetic (deterministic)	csv, yfinance ([market])	FINSIGHT_DATA_BACKEND

Optional deps are imported lazily and degrade gracefully: if FinBERT isn't installed, FinSight warns and falls back to the offline lexical backend rather than crashing.

Why this is honest

• No look-ahead. Positions are weights.shift(1) — a position can never see the return that justifies it (enforced by a test).
• Costs are real. Turnover is charged at 10 bps per unit; the high-turnover raw signal is deliberately not profitable until smoothing controls turnover.
• The hard cases stay in. The synthetic lexical model recovers sentiment at ~0.84 corr, not 1.0 — the gap is exactly what a stronger model (FinBERT) would close. The signal IC (~0.05) is in the realistic range for daily news alpha, not a fantasy.

See docs/ARCHITECTURE.md and docs/DECISIONS.md for the design write-up.

License

MIT — see LICENSE.