Forecasts

A forecast says what comparable past events did to travel: how often they led to website outages, check-in problems or cancellations, with an 80% interval and n, the number of events behind each number. Analog forecasts exist for IT outages today; other signal types get recommended actions and airport base rates.

Where do forecasts appear?#

  • GET /v1/signals/{id} includes impact_forecast for any signal.
  • GET /v1/forecast?provider=...&failure_layer=... is a what-if: the forecast for an outage that has not happened. failure_layer is required; GET /v1/failure-layers lists the values.
  • Exposure matches on IT outages carry a short forecast: the top four outcomes, recommended actions and expected duration.
  • Webhook payloads for IT outage signals include impact_forecast.
Request
curl "https://app.disruptapi.com/v1/forecast?provider=cloudflare&failure_layer=edge_cdn_dns" \
  -H "Authorization: Bearer $DISRUPT_KEY"
Response (illustrative values, shortened)JSON
{
  "data": {
    "basis": "analogs",
    "failure_layer": "edge_cdn_dns",
    "failure_layer_label": "Edge, CDN and DNS",
    "analog_count": 14,
    "analogs": [
      {
        "id": "2025-11-cloudflare",
        "title": "Cloudflare outage (bot management file exceeded size limit)",
        "provider": "cloudflare",
        "failure_layer": "edge_cdn_dns",
        "starts_at": "2025-11-18T11:20:00Z",
        "duration_h": null,
        "outcomes": { "website_or_app_down": 1 },
        "url": null
      }
    ],
    "outcomes": [
      {
        "outcome": "website_or_app_down",
        "label": "Airline websites or apps down",
        "probability": 0.778,
        "interval": [0.594, 0.931],
        "interval_level": 0.8,
        "k": 6,
        "n": 7,
        "analog_level": "same failure layer, any provider",
        "low_confidence": false,
        "evidence": "6 of 7 analogs (documented); e.g. Cloudflare outage (bot management file exceeded size limit)",
        "base_rate": null,
        "lift": null,
        "above_base_rate": null
      },
      {
        "outcome": "carrier_cancellation_spike",
        "label": "A US carrier's own cancellations jump (BTS, carrier-caused)",
        "probability": 0.312,
        "interval": [0.172, 0.464],
        "interval_level": 0.8,
        "k": 4,
        "n": 14,
        "analog_level": "broader group of IT outages",
        "low_confidence": false,
        "evidence": "4 of 14 analogs (bts); ordinary days show this 21% of the time, so these events do not clearly raise it",
        "base_rate": { "probability": 0.21, "windows": 56 },
        "lift": 0.102,
        "above_base_rate": false
      }
    ],
    "expected_duration_hours": {
      "p50": 2.4,
      "p90": 7.9,
      "n": 11,
      "analog_level": "same failure layer, any provider"
    },
    "carriers_at_elevated_risk": [
      {
        "carrier": "AS",
        "confidence": "medium",
        "reasons": [
          "alaskaair.com uses cloudflare (edge_cdn)",
          "alaskaair.com uses cloudflare (dns)"
        ],
        "dependency_layers": ["dns", "edge_cdn"],
        "past_events": 0
      }
    ],
    "contrast": "Different failure layer from CrowdStrike (July 2024), where an endpoint agent crashed Windows machines at airlines and airports and Delta cancelled about 7,000 flights. A edge, cdn and dns failure usually leaves those machines running.",
    "recommended_actions": [
      "Airline and booking websites may fail; check in at the airport or by app if the web is down",
      "Allow extra time at airports for groups that cannot check in online"
    ],
    "notes": [
      "Probabilities use a Beta-binomial model with a uniform prior; n is the number of past events where the outcome was measured or documented. Treat any outcome marked low_confidence (n under 5) as indicative."
    ]
  }
}

Which outcomes are forecast?#

OutcomeMeaning
website_or_app_downAirline websites or apps down
checkin_disruptedCheck-in disrupted
any_cancellationsFlight cancellations at one or more carriers
carrier_cancellation_spikeA US carrier's own cancellations jump (BTS, carrier-caused)
us_cancellation_spikeUS-wide carrier-caused cancellations double
ground_stops_above_normalFAA ground stops well above normal
passenger_comms_disruptedPassenger notifications (email, SMS) delayed or failing

Outcomes with no history for the provider and layer are left out rather than guessed.

How are analogs chosen?#

For each outcome separately, DisruptAPI looks for past events in the catalogue (GET /v1/events) at the most specific level that has at least 5 events where that outcome was measured or documented:

  1. Same provider and failure layer, e.g. Cloudflare edge failures.
  2. Same failure layer, any provider, e.g. every edge, CDN and DNS failure.
  3. Broader group: all IT outages that matter for travel, or all cloud outages for the cloud layers. At this level only measured outcomes (BTS and FAA data) are pooled, because documented outcomes come from notable events that lean towards severe cases.

If no level reaches 5, the level with the most cases is used and the outcome is marked low_confidence. analog_level says which level was used.

How is the probability calculated?#

A Beta-binomial model with a uniform prior. With k of n analogs showing the outcome, the probability is (k + 1) / (n + 2) and the interval is the central 80% of the Beta(k + 1, n − k + 1) distribution. Two of nine gives 0.273 with an interval of 0.116 to 0.45. The interval is wide when n is small, on purpose.

What are base rates and lift?#

Some outcomes happen on ordinary days too. US carrier-caused cancellation spikes, for example, show up on roughly one ordinary day in five. For measured outcomes, DisruptAPI checks the same outcome on matched control windows of the same length with no catalogued event, and reports:

  • base_rate.probability: the share of control windows with the outcome, and base_rate.windows, how many were checked.
  • lift: the forecast probability minus the base rate.
  • above_base_rate: true only when the whole 80% interval sits above the base rate. That is the test for "this kind of event makes the outcome more likely".

Documented outcomes (from reports about notable events) have no control windows, so their base rate fields are null.

How should I read a forecast?#

  • Lead with outcomes where above_base_rate is true. They are listed first. An outcome with above_base_rate: false happens about as often without the event; its evidence text says so.
  • Always show n, and say "indicative" when low_confidence is true.
  • Read the interval, not only the point. 0.78 with an interval of 0.59 to 0.93 from 7 events is a strong signal; 0.40 from 3 events is not.
  • Use carriers_at_elevated_risk to decide which airlines to watch. high needs both a dependency and a past impact; dependencies alone give low or medium.

Outcome fields#

FieldTypeDescription
outcomestringOutcome id, from the table above.
labelstringWhat the outcome means, in words.
probabilitynumberPosterior mean, 0 to 1: (k + 1) / (n + 2).
interval[number, number]Central 80% credible interval.
interval_levelnumberAlways 0.8.
kintegerPast events where the outcome happened.
nintegerPast events where the outcome was measured or documented. The sample size; always report it.
analog_levelstringWhich group of past events was used: "same provider and failure layer", "same failure layer, any provider" or "broader group of IT outages".
low_confidencebooleanTrue when n is under 5.
evidencestringPlain-words summary: "2 of 9 analogs (bts, documented); e.g. ...", and a note when ordinary days show the outcome as often.
base_rateobject | null{probability, windows}: how often the outcome happens in matched ordinary windows with no catalogued event. Null for documented-only outcomes.
liftnumber | nullprobability minus base_rate.probability. Null without a base rate.
above_base_rateboolean | nullTrue only when the whole interval is above the base rate. Null without a base rate.

Forecast fields#

FieldTypeDescription
basisstringanalogs for IT outages; insufficient_history for other signal types.
failure_layerstring | nullThe layer the forecast is for.
failure_layer_labelstring | nullIts label.
analog_countintegerDistinct past events used across all outcomes and the duration estimate.
analogsobject[]Up to 6 named comparable events: id, title, provider, failure_layer, starts_at, duration_h, outcomes (outcome id to 0 or 1), url. Landmark events first.
outcomesobject[]One entry per outcome with any history, fields below. Outcomes that do not beat their base rate are listed last.
expected_duration_hoursobject | nullp50, p90, n and analog_level from past durations. Null when none are known.
carriers_at_elevated_riskobject[]Up to 25 airlines: carrier, confidence (low, medium, high), reasons, dependency_layers, past_events.
contraststring | nullHow this layer differs from an endpoint crash like CrowdStrike in July 2024.
recommended_actionsstring[]Practical steps for this failure layer or signal type.
notesstring[]Method notes to show next to the numbers.
airport_base_ratesobjectNon-IT signals with airports: per airport, that month's cancellation and delay base rates where measured (US airports).

What are the caveats?#

  • US-centred measurements. BTS cancellation outcomes cover US carriers only. Outside the US, outcomes come from documented reports, and history is thin.
  • Non-IT signals have no analogs yet. Weather, hazards, advisories and others return basis: "insufficient_history" with recommended actions, plus monthly airport base rates (airport_base_rates) where they exist, which is at US airports.
  • Selection bias in documented events. Famous outages are better documented than small ones. They only count within the same failure layer.
  • The catalogue grows. Numbers change as new events and outcomes are added. Do not cache a forecast for long, and do not compare numbers across months as if the method were fixed.

Last updated 2026-11-18. Questions or a mistake on this page? Contact us.