SENECArisk intelligence ← Back to seneca-risk.com

Methodology

Last updated: 30 June 2026

Seneca estimates how much the S&P 500 could lose — the depth of a potential drawdown — over the next 30, 90, 180 and 365 days. It is a continuously-updated risk gauge built on three decades of market history. This page explains how it works and, just as importantly, where it is weak. We would rather you trust the honest version.

1. What Seneca forecasts — and what it doesn't

Seneca forecasts the magnitude of forward drawdown risk, not the timing of a specific crash and not the direction of tomorrow's return. It answers “how much downside is the current environment carrying?” — a risk dial, not a trade signal. A high reading means conditions resemble past pre-drawdown regimes; it is not a promise that a decline is imminent, and a low reading is not a guarantee of calm.

2. The data

Each forecast is computed from a rolling window of 85 features per observation, drawn from public sources going back to 1986:

  • Markets & valuation — S&P 500 price action, volatility and momentum structure, and long-run valuation context.
  • Macro & credit — a broad panel of economic series (rates, the yield curve, credit spreads, inflation, housing, labour, consumer and industrial activity).
  • Stress & sentiment — geopolitical-risk and global event signals.
  • Policy text — Federal Reserve communications (FOMC statements and the Beige Book), scored for tone with a language model.

3. No look-ahead — strict anti-leakage

The single most common way backtests lie is by letting future information leak into past predictions. Seneca enforces against this at every layer: every feature for a given date uses only data available on or before that date; normalisation statistics are fit on the training portion only; labels are derived strictly from future price paths; and the evaluation harness asserts that the training window always ends before the date being predicted. No model ever sees the answer it is being scored on.

4. Walk-forward validation — the only number we trust

Every figure Seneca publishes comes from walk-forward evaluation, not an in-sample fit. Starting in 1996, the models are trained on all history up to a point in time, asked to predict forward, then advanced one week and retrained — roughly 1,600 times across the record. This reproduces how the system actually runs in production: it only ever knew what was knowable at the time. The full public record holds more than 11,000 such out-of-sample predictions.

5. Three models in consensus

No single model is reliable across every regime, so Seneca combines three complementary families and reports their consensus:

  • A regularised linear model that maps the current feature state to a risk level — the most stable component across regimes.
  • A historical-analog retrieval model that finds the moments in history most similar to today and looks at what followed.
  • A neural embedding model that learns a similarity space tuned to drawdown outcomes.

The published magnitude is the consensus of the three; the spread between them is itself informative and is surfaced as an uncertainty band. When the models disagree, Seneca tells you.

6. How to read the numbers honestly

Headline correlation can flatter a forecaster at long horizons, because drawdown levels are highly autocorrelated: a model that simply echoes “tomorrow looks like today” already scores well. So the metric we hold ourselves to is skill above that naive persistence baseline: how much the model improves on just guessing the recent level. On that harder test the system adds meaningful, positive skill, most at shorter horizons, and we report it rather than the inflated raw number.

7. What Seneca cannot do

Seneca is probabilistic and can be wrong. It is a statistical estimate of risk, not a forecast of certainty. Specific limits to keep in mind:

  • It estimates drawdown magnitude, not exact timing. An elevated reading can persist before a decline, or without one.
  • Its consensus edge depends on continuous retraining. On a frozen model with no updating, only the linear component retains dependable skill; the analog and neural components need fresh data to stay calibrated. Seneca retrains weekly to keep that edge live.
  • Long-horizon accuracy is partly the autocorrelation effect described in §6; read the short-horizon, persistence-adjusted skill as the honest signal.
  • Past out-of-sample performance does not guarantee future accuracy. Markets change; a model trained on the past can be surprised by the genuinely new.

8. Reproducibility & review

The methodology — feature engineering, the walk-forward design, the anti-leakage proofs and the ensemble construction — is documented in a master's research thesis that received strong positive feedback from its reviewing scientific committee. Every claim on this site traces back to a published walk-forward result, and the live dashboard reports rolling recent accuracy alongside all-time figures so you can watch the model's health in real time.

9. This is not investment advice

Seneca is provided for informational and educational purposes only and is not financial, investment, legal or tax advice, and not a recommendation to buy or sell any security. You are responsible for your own decisions; consult a licensed professional before acting. See our Terms for the full disclaimer.

PrivacyTermsCookies Cookie settings