Max Agent
Articles

Give Every Alert a Review: Usefulness, Missed Events and Attention Cost

Build an alert review around a defined task, an independent event register and a time budget. Our proposed workflow includes explicit rules for changing or retiring notifications.

MaxData EditorialPublished Updated

Start an alert review with a question you can answer from a log: “Did this notification help me complete the check I had planned?” For this article, we propose an editorial workflow for evaluating cryptocurrency notifications without turning the exercise into a trading recommendation. Choose the information task, record the evidence and decide whether the alert deserves another observation period. Treat every score, threshold and decision rule below as our proposed convention, rather than an established performance standard.

Define the job before scoring the alert

The MaxData alert guide describes a basic Alert as a notification triggered by one condition and a Pro Alert as a combination of conditions. It also describes a time window, a period of inactivity after triggering and data choices such as open, close, minimum or maximum. These are statements about the documentation, without confirming availability for a particular account. Pro Alerts guide.

Our proposed specification begins with the check you intend to perform. For example: “When the selected metric crosses my recorded level on a completed candle, review the underlying chart within my chosen response window.” Keep the task educational: inspecting data, checking an announcement or recording an observation. Do not make placing a trade a requirement for usefulness.

Create a versioned record containing the instrument, market, metric definition, unit, data provider, candle interval and condition. Name the exact price or metric field being evaluated. Add the observation dates, the intended response deadline and the later analysis horizon. If you choose a fifteen-minute candle and a four-hour follow-up, label both as your test assumptions. Give any altered threshold a new version number.

Build two registers with different jobs

Our proposed notification register contains one row for every delivered message. Include the alert version, event identifier, delivery time, review time, classification and minutes spent. Keep copies sent through different channels visible, while linking them to the same event when appropriate. Decide before testing whether repeated messages count as reminders you requested or unwanted duplicates.

The second register is your proposed reference list of events. Construct it through a separate examination of the selected data over a declared period. For a manageable first exercise, choose one instrument and inspect every completed candle in that period against your fixed event definition. Record the source, timestamp, observed value and any matching notification identifier.

Do not describe a limited inspection as complete coverage. If you examine only selected hours, restrict the review conclusion to those hours. Mark gaps in reference data as unassessed. Where the event cannot be reconstructed, leave the match unresolved. In our workflow, an unresolved row is a request for evidence, rather than an automatic failure or success.

Our proposed record design

Messages and events need separate rows

Notification register
One row
One delivered message
Record
Delivery, classification, minutes
Link
Matching event identifier
Reference register
One row
One defined event
Record
Source, timestamp, observed value
Link
Matching message identifier
Link the two registers with identifiers and preserve unresolved matches. A proposed comparison of notification records and independently reviewed event records.

Use a scorecard with explicit denominators

We propose defining a useful notification as one that matches the event definition, arrives within the chosen deadline and leads to the planned check. A documented decision to take no further action qualifies under this convention. A message that arrived too late does not qualify, even if you subsequently reviewed it.

For the notification score, divide useful messages by all delivered messages for that version and observation period. Display the counts beside the ratio, including unresolved messages. Assign each message one main classification using a recorded order: first event validity, then delivery deadline, then duplication. Keep additional diagnostic notes outside that main classification.

For the missed-event score, divide reference events without a timely matching notification by all assessed reference events. Treat this as a separate view of the same review. Under our proposed convention, a late message can correspond to a missed event; do not add both counts and call the result a total number of distinct incidents. If a denominator is zero, report the score as undefined and explain which observations are absent.

Measure the time the alert asks you to spend

Our proposed attention log separates reading, source checking, chart inspection and writing the conclusion. Record minutes for each activity, then report the total review time and the time per useful notification. When there are no useful notifications, show the total minutes and leave the per-useful figure undefined. Add subscription or delivery charges in a separate field using a stated allocation rule.

For timing, collect four timestamps where available: the event according to your definition, the point at which its data became available, message delivery and the start of your review. Name both endpoints of every reported delay. Use “unknown” when a timestamp is unavailable; do not quietly substitute the moment you opened the message.

We propose reporting the median observed event-to-delivery delay alongside its sample count and measurement period. Define “typical delay” this way in your own report, rather than borrowing a product description. Choose the acceptable deadline before the observation period and preserve late deliveries in the record. Keep your response time separate from the delivery measurement.

Our proposed measurement sequence

Name the endpoints

  1. 1

    Event

    Record the time specified by your event definition.

  2. 2

    Data available

    Record availability for evaluation when observable.

  3. 3

    Delivered

    Record receipt through the selected channel.

  4. 4

    Review started

    Record when you began the planned check.

Leave unavailable timestamps unknown and record review minutes separately. Four proposed timestamp fields from the defined event through the start of the reader's review.

Give historical testing a limited assignment

The English MaxData backtest guide says a backtest can report how many times an Alert or Pro Alert would have triggered in the past. It describes average and median price-change lines before and after triggering, and testing either a single condition or several conditions together. Backtests guide.

In our proposed workflow, use that historical view to draft a hypothesis and estimate a workable observation schedule. Then evaluate a frozen version during a separately designated period. Label historical observations as historical. Require another recorded assessment before making any statement about a later period; do not annualize a past trigger count into an expectation within this review.

Keep an experiment ledger listing every threshold you tried, the period used to select it and the period reserved for evaluation. As our check for fitting choices to history, ask whether you selected the configuration after inspecting its result. If so, label that portion of the work as design. Do not combine it silently with the later evaluation sample.

If you add simulated trades, create a separate exercise with explicit fees, slippage, order size, liquidity assumptions and execution rules. Require those fields before discussing simulated outcomes. Keep notification usefulness and simulated trading results in separate sections, with separate denominators and limitations. This article’s proposed scorecard does not assign a return target.

Diagnose a weak result before changing the rule

Use our proposed diagnostic sequence whenever the review looks disappointing. First verify that the reference events and notifications use the same instrument, provider, field and interval. Next inspect delivery timing. Then examine duplicates and your own review time. Finally check whether your definition of usefulness changed during the observation period.

For each alternative explanation, write down the evidence needed to decide it. An unexplained timing discrepancy should lead to a timestamp check. A disputed event match should lead to inspection of the recorded data. A suspected duplicate should lead to comparison of event identifiers. Leave the diagnosis open when that evidence is unavailable.

We propose invalidating the combined score when versions were mixed, the data source changed without documentation or reference coverage cannot be established. Start a new assessment period after repairing the record. If there are too few assessed events for your preselected requirement, use the decision “insufficient evidence”, with a deadline and an explicit limit on further observation.

Write retirement rules while the test is still unscored

Before observing results, choose your own attention budget, tolerated misses, minimum assessed sample and review date. These are proposed operating limits for your workflow, without a universal numerical recommendation. Give each alert an owner responsible for the decision. Include a condition under which its information task is no longer relevant to your process.

Use three main decisions: retain the frozen version, revise one element or retire it. A revision starts a new version and a new evaluation record. Retirement should preserve the specification, evidence and reason, together with the conditions required for reconsideration. For any automation you have connected, add a dependency check and confirmation that associated actions have been disabled before closing the retirement task.

Finish with a short review note: task, version, observation scope, useful-message count, missed-event count, attention minutes and decision. List unresolved questions and conditions that would invalidate the conclusion. Make the next review date explicit if you retain the alert.

The official MaxData Help Hub presents an assistant for questions about MaxData and the workspace. MaxData Help Hub. Our proposed support request includes the alert version, expected event, observed timestamps and one precise question about the discrepancy. Use the answer to complete your evidence record before assigning a final diagnosis.

Our proposed review cycle

Decide with a preserved record

  1. Set limits

    Choose the sample requirement, attention budget and review date.

  2. Observe

    Maintain both registers for one frozen version.

  3. Assess

    Check coverage, timing and unresolved explanations.

  4. Record decision

    Retain, revise or retire; defer explicitly if evidence is insufficient.

Set your own limits before scoring; keep the evidence when a version is retired. Proposed stages of setting limits, observing a frozen alert, reviewing evidence and recording a decision.

Sources

  1. MaxData Help Hub – Pomoc i przewodniki Mój workspace — agent.maxdata.app; accessed November 20, 2026
  2. Pro Alerts — agent.maxdata.app; accessed November 20, 2026
  3. Backtests — agent.maxdata.app; accessed November 20, 2026

Start working with your own agent

The agent knows this article context and still follows the assigned department checklist.

Create a MaxData account
Max Agent