AstraEye Labs
Sign InEnter the Lab
AstraEye Labs · Research methods · ATLAS
ATLASU.S. equitiesFreelancer plan

Volatility Outlook

Forecast how much each stock and fund will move over the next 5 and 21 sessions, with any scheduled earnings report priced in from the company's own past reactions

  1. 01
    What it answers
  2. 02
    How it is computed
  3. 03
    Tested on past data
  4. !
    Keep in mind
  5. →
    Read a real report
Impersonal research. AstraEye Labs is not a registered investment adviser, does not give financial advice, and does not know your goals or risk tolerance. This research does not account for your individual goals, risk tolerance, financial situation, or other circumstances. Past performance does not predict future results.
01The analysis

What it answers

Answers how much each stock and fund in scope is expected to move over the next 5 and 21 trading sessions: a forecast of its annualized realized volatility, computed by code from up to ten years of adjusted daily prices (Tiingo) and the times the SEC accepted each company's earnings 8-K filings, with no options data and no AI model. One model, pooled across every stock and fund AstraEye holds ten years of prices for and refitted once a quarter, forecasts ordinary volatility from the asset's recent daily price ranges against its own three-year level and the market's. An earnings report projected inside the window from the company's release a year earlier is priced from the company's own past report reactions, shrunk toward its sector's. For each asset it states the forecast, its ordinary part, any projected report reaction with its share of the forecast, and the realized volatility of the last 5 and 21 sessions beside it, all from the latest completed close and the same for every reader. Funds are forecast from their own prices, funds that hold futures contracts included. It was tested once on past data, 2020 to 2026, under a plan written before any figure was computed, against two standard baselines; its results there, computed after the fact, are shown with their limits beside this method on AstraEye's methods page, and past results do not guarantee or predict future results. A forward test is running, judged on 17 May 2027, and its result will be published that day whatever it shows. The forecasts may be wrong, state no direction and name no action.

Scientist
ATLAS
Asset class
U.S. equities
Securities it can describe
Stocks and ETFs
Plan
Included from the Freelancer plan
02The method

How it is computed

Forecasting volatility from prices alone: how much a price is likely to move, never which way.

REALIZED VOLATILITY. How much prices actually moved, measured from each session's open, high, low and close (the Parkinson, Garman-Klass and Rogers-Satchell range estimators, plus the overnight gap) and annualized by about 252 sessions a year. A forecast of it is a forecast of the size of moves, not their direction.

VOLATILITY CLUSTERS. Large moves tend to follow large moves, and calm follows calm, so recent volatility says something about the next weeks. The HAR model (Corsi, 2009) reads it at several speeds at once: the last session, week, month and quarter. Working in logs, relative to each series' own three-year level, lets one model serve a quiet utility and a volatile chip maker alike.

POOLING. One stock's history is short and noisy. Fitting one model across hundreds of stocks and funds, as Bollerslev, Hood, Huss and Pedersen (2018) do, gives far steadier coefficients than one stock's own; the asset's level and recent readings still set its forecast. The market's last month (SPY's) and the share of recent variance carried by falls enter too.

SCHEDULED REPORTS. An earnings report is the one large move a calendar can see coming. Sessions that reacted to a report are set apart from ordinary volatility, and a report projected inside the window (the company's release a year earlier, moved 364 days forward to the same weekday) is priced from the variance of the company's last twelve report reactions, shrunk toward its sector's average when it has few. The forecast blends the two by the number of report sessions in the window.

TESTED BEFORE IT WAS SHOWN. The method, its two baselines (the trailing realized volatility and an exponentially weighted average) and its pass standards were written down before any figure was computed, and it was run once on 2020 to 2026, each forecast made only from data available on its date. Its results on that past data, computed after the fact, are shown with their limits beside this method on AstraEye's methods page; they are results on historical data, not a record of forecasts published at the time, and past results do not guarantee or predict future results.

THE FORWARD TEST. Every week AstraEye records the forecast for every stock and fund it covers before the next session opens. The test is judged once, on 17 May 2027, against standards declared on 12 November 2026, and its result will be published that day whatever it shows.

Keep in mind

WHAT IT IS NOT. Not a forecast of direction, and it names no action. It reads no options data or implied volatility, so it cannot see what option buyers expect, and a report date the company has announced or moved, or an unscheduled event, is not known to it. The 364-day projection, the twelve-reaction window and the shrinkage weight are our own convention.

03Past data

Tested on past data

On past data, 2020 to 2026: AstraEye's study of 12 November 2026 tested this method once, under a plan written before any figure was computed, on weekly forecasts from January 2020 to October 2026 for every stock and fund it holds ten years of prices for, each forecast made only from data available on its date. Each figure below is how much lower the method's QLIKE loss (the usual loss for variance forecasts; lower is closer) was than a baseline's.

21 sessions, against the trailing 21-session realized volatility
24.0% lower (197,947 forecasts)
21 sessions, against an exponentially weighted average (RiskMetrics, lambda 0.94)
19.3% lower (197,947 forecasts)
5 sessions, against the trailing 5-session realized volatility
38.4% lower (199,700 forecasts)
5 sessions, against the exponentially weighted average
24.8% lower (199,700 forecasts)
21-session windows holding an earnings report, against the trailing baseline
57.2% lower (52,662 forecasts)

Corrected on 12 November 2026, after the study ran: some price files held days with no trading, before a company listed or while it was an empty shell company, each a quote carried over from an earlier day with no volume. The study had read them as calm days. The figures above leave out the forecasts whose own sessions held such a day, 65 at 5 sessions and 94 at 21 sessions. Before the correction they were 42.2% and 24.6% lower at 5 sessions and 23.7% and 19.1% lower at 21 sessions. The method now counts such a day as missing.

  • The study's universe is today's stocks and funds, read back to 2020: a company delisted or dropped from AstraEye's coverage since then is not in it, which can make past results look better than a test on the universe of each date would.
  • In 21-session windows without a report, and for funds alone, the difference from the trailing baseline could not be told from zero (6.5% and 8.3% lower).
  • In 2021, the year after a crisis, its 21-session loss was higher than both baselines (0.20 against 0.17 and 0.14): its three-year anchor held forecasts toward a level the market had left.
  • At 63 sessions the results could not be told from the baselines', it did worse than both in 2021, 2022 and 2023, and its squared error was larger than the trailing baseline's, so the 63-session forecast is not shown.
  • These are results on historical data, not a record of forecasts published at the time. Past results do not guarantee or predict future results.

These results are based on simulated or hypothetical performance results that have certain inherent limitations. Unlike the results shown in an actual performance record, these results do not represent actual trading. Also, because these trades have not actually been executed, these results may have under-or over-compensated for the impact, if any, of certain market factors, such as lack of liquidity. Simulated or hypothetical trading programs in general are also subject to the fact that they are designed with the benefit of hindsight. No representation is being made that any account will or is likely to achieve profits or losses similar to these being shown.

What was tested: the Volatility Outlook's forecasts of realized volatility over 5, 21 and 63 sessions, for every stock and fund AstraEye holds ten years of prices for, weekly from January 2020 to October 2026, computed on past data and compared with the volatility that followed. The comparisons with both baselines were declared before the study ran. The figures measure forecasts: no trade was simulated or executed.

✦Read a real report

From a method to a finding