How does Hesper Atlas generate buy, hold and sell signals?

Hesper Atlas routes each asset to a rules-based trend engine suited to its behavior, then converts price trend, momentum, volatility and risk controls into a buy, hold, reduce or sell state.

What does held-out backtesting mean?

Hesper Atlas tunes on the early 60% of each instrument and scores the later 40%. That tail was originally withheld, but subsequent research reused it, so it is now described as validation rather than a pristine unseen test. It reduces one form of curve-fitting risk but cannot eliminate it or guarantee future results.

How well does the Hesper Atlas engine actually perform?

On the later 40% static-tail validation across all 152 evaluated names (July 2026, after costs, next-day fills), the engine's median worst drawdown was -30% versus -57% for buy-and-hold, 150 of 152 names had a shallower worst drop, and the median return was +22.7% per year versus +18.9% holding. Later research reused the tail, so it is not a pristine unseen test. These are backtests, not live results; the public replay and forward ledger are separate artifacts.

Are Hesper Atlas signals financial advice?

No. Hesper Atlas is an educational tool and impersonal publication. Signals are generated algorithmically and are not personalized recommendations.

How Hesper Atlas creates its signals

The methodology behind the buy, hold and sell calls, the backtests, and the limits of what the system can know.

Version: ha-methodology-1.0.0 · September 1, 2026 . Historical replay . Replay JSON . Validation JSON . Walk-forward JSON . Version manifest . Forward commitments . Agent verification guide

Short answer: Hesper Atlas is a rules-based trend system. It classifies each asset by behavior, routes it to a matching engine, and turns end-of-day price data into a buy / hold / reduce / sell state. Four evidence layers stay separate: legacy static-tail validation, cache-frozen rolling walk-forward validation, a retrospective replay of today’s rules, and an append-only forward publication ledger that starts at deployment. None is a promise of future returns.

1The signal stack

Asset routingDifferent assets behave differently. Indexes, crypto, low-volatility compounders and high-volatility growth names are not forced through one formula.

Trend stateThe engine asks whether the asset is in a durable uptrend, a weakening trend, a fresh risk-off state, or a confirmed exit zone.

Trade planThe app converts the signal into exact levels to own, add, trim, reduce or sell. Free covers three names a day, full trade plan included; Pro opens the whole universe.

Historical replayToday’s rules are replayed over up to five years of end-of-day data, with winners, losses and open replay positions. Retrospective, not live performance.

2What the engines look at

The engines use end-of-day market data: price trend, momentum, volatility, moving-average structure, drawdown risk and risk-control rules. The system is deliberately impersonal: everyone looking at the same asset on the same day sees the same signal.

The app also shows contextual layers such as market heat, macro regime, business quality and valuation risk. Those help explain the environment around a signal, but the signal itself remains systematic.

3What static-tail validation means

The evaluation splits each instrument chronologically: the early 60% is the tuning span and the later 40% is the scoring span, after costs with next-day fills. The later tail was originally withheld from the early fit. Subsequent research and campaign selection reused those dates, however, so it is no longer a pristine unseen test set. Hesper Atlas now labels it validation evidence and publishes that limitation in the machine-readable run manifest.

This split helps expose some curve fitting but cannot eliminate it. Market structure can change, data can be wrong, the curated universe carries selection and survivorship bias, and future results can diverge sharply from backtests. The cached Yahoo earnings-event index can lag the present by roughly a year; unknown quarters leave those overlays inert rather than filling them with later knowledge. Rolling walk-forward audits are the stricter promotion authority.

When Hesper Atlas cites a result, it should be read as historical or backtested evidence under the stated assumptions, not as an investment forecast.

4What the results honestly show

The headline evidence is the rolling walk-forward: each January the rules are chosen on data available to that date, then traded for the next calendar year, 2016 to 2026, across the 144 names with enough history, next-open fills, after costs. The untuned class engine returned a median +18.5% a year versus +20.7% buying and holding, with a median worst drawdown of −50% versus −57%, shallower on 128 of 144 names and an equal Sharpe ratio. The drawdown reduction is the broad part of the result. It comes from stepping aside in confirmed downtrends and appears across most evaluated names; the return cost concentrates in sharp recovery years.

The per-name tuned rules did worse: a median 14.8% a year in the same test, and every regularised selector, plus class-level pooled selection with the traded name excluded from its own vote, converged to the untuned engine from below. Per-name tuning carries no out-of-sample information, and the self-audit publishes the full result. The older static-tail validation (+22.7% versus +18.9%, −30% versus −57% on 152 names, July 2026) used rules chosen with that tail in view; it remains public as validation JSON but is no longer the headline. Read every return figure with care: the universe is itself selected for engine fit.

We run the stricter audits most signal products skip. In the yearly expanding-window walk-forward test, the adaptive selector frequently lagged buy-and-hold and the champion-free class base on return, while generally reducing drawdown. Those weak names and years are published in the row-level artifact. A deflated-Sharpe significance audit likewise finds that most per-name return edges are not statistically distinguishable from luck once the number of tried variants is counted. That is why Hesper Atlas does not market the optimizer as a return machine, and why the separate historical replay includes winning, losing and open signals under today’s rules instead of showing a highlight reel.

5Where to verify claims

Historical replay: the human-readable current-rule replay with summary statistics, every closed replay row, open positions and caveats.
Replay JSON: the machine-readable source with data type, timestamps, provenance, closed and open replay signals and caveats.
Validation JSON: the content-derived run id, SHA-256 artifact digest, assumptions, limitations, aggregate summary and all 152 per-name rows.
Rolling walk-forward JSON: hashes for the isolated price/earnings snapshot (rechecked against the frozen files at publication), the exact evaluation-component hash list, annual fit-through and out-of-sample dates, candidate choices, weak periods and per-name comparisons.
Pooled selection JSON: the class-pooled walk-forward audit, with every class-year vote, leave-one-out and strict variants, stitched per-name metrics and the frozen-input hash; the write-up is at /research/per-name-tuning.
Methodology and calculation manifest: the stable methodology version, content-derived published-decision/evaluation-pipeline versions, component hashes and the explicit limit that these are first-party integrity hashes, not an independent timestamp.
Forward commitments: append-only deployment-onward publication ids and SHA-256 commitments; exact frozen rows are available to subscriber agents.
Hesper Atlas for Agents: the verification workflow, copyable prompts, data-status model and current MCP limitations.
Terms of Service: the formal caveat that Hesper Atlas is educational, not personalized investment advice.
LLM answer context: a compact source map for AI answer engines.

6What Hesper Atlas is not

Hesper Atlas is not an investment adviser, broker or fiduciary. It does not know your portfolio, tax situation, risk tolerance or time horizon. Its output is an educational, algorithmic market-analysis publication, not a personalized recommendation to buy, sell or hold any security.