Max Agent
Articles

From backtest to live observation: a review process for trading automation

Build an evidence file with a reserved test period, walk-forward reviews, paper trading logs, and explicit conditions for allowing or stopping automation.

MaxData Editorial TeamPublished Updated

Before allowing a rule to trigger automated actions, write down what evidence you want to review. This article proposes an editorial workflow for doing that: specify the experiment, reserve an evaluation period, repeat the review through time, and observe new events under a documented execution model. The example schedules and approval criteria are our proposed conventions, not validated thresholds for trading success.

The English MaxData backtest guide describes checking historical alert frequency and displaying average and median price changes around alert triggers. Backtests guide. We use that documented scope as a starting point for an evidence file. This is a methodology exercise, with no recommendation to trade a particular asset or direction.

Write the experiment before choosing the winner

Our proposed first deliverable is an experiment card. Give it a question, an owner, and a version identifier. For example: “What price changes would we record during the four hours after this precisely defined condition?” Treat the four-hour horizon as an exercise assumption, not a suggested holding period.

Require the exact instrument, trading venue, market type, quote currency, price source, and date range. Add a separate definition and source for every input metric. Keep the data interval separate from the rule evaluation schedule: in our example, use fifteen-minute candles and evaluate after each candle closes.

Include fields for missing observations, time zone, and typical data delay. Leave the delay marked unknown until measurements are available. Our proposed measurement is the elapsed time between a data timestamp and observed availability. Record the measurement period and sample count alongside the resulting summary. Finally, specify the observation horizon and the assumptions under which the experiment should be considered invalid. Do not approve an incomplete card under this checklist.

Reserve a period and freeze the rules

For this workflow, we use in-sample to mean the historical segment assigned to rule development. We reserve out-of-sample for a later segment evaluated after the rule has been frozen. These are the working conventions of our proposed protocol.

Choose the date boundary before examining evaluation results. As an illustrative schedule, assign six months to development and two subsequent months to evaluation. Set a rule for observations whose measurement horizon crosses that boundary. The example does not establish an adequate sample size; write your own evidence requirement before running the assessment.

The Pro Alerts documentation describes a condition window, a sleep period after triggering, and a choice of input values including open and close. Pro Alerts guide. In our specification, require those settings to be written down alongside entry, exit, expiry, and repeat-trigger rules.

Freeze the complete specification under a version number. Our proposed convention is that changing any threshold or execution assumption creates a new version. Keep a register of every attempted variant and its selection rationale. For the review of possible fitting to history, ask how many candidates were inspected and whether the selection rule was written before the results were seen.

Proposed research protocol

Prepare the reserved evaluation

  1. 1

    Choose dates

    Assign development and evaluation segments before inspecting results.

  2. 2

    Specify behaviour

    Record triggers, exits, expiry, and execution assumptions.

  3. 3

    Freeze a version

    Give the complete specification an identifier.

  4. 4

    Review evidence

    Apply the previously written evaluation requirements.

An editorial sequence for documenting the experiment; these steps do not establish trading effectiveness. Four proposed preparation steps: choose dates, specify the rule, freeze a version, and evaluate the reserved segment.

Read the reserved test as an evidence file

Our proposed review separates event observations from a simulated account ledger. The English guide describes its backtest output as average and median price-change lines before and after an alert, and says a backtest can cover one condition or several together. Documented output. Use the documented description when naming the report you have, then record any additional modelling as a separate exercise.

For each evaluation period, our checklist requests the number of triggers, completed observations, exclusions, assumed costs, and unresolved cases. For an account simulation, also request a ledger and a stated definition of drawdown. Define the balance calculation and treatment of overlapping events before calculating a summary.

Assign one of three editorial statuses: sufficient evidence for the next review, insufficient evidence, or specification failure. Do not choose a numerical minimum after seeing the trigger count. If your prewritten requirement is unmet, leave the status unresolved. Add alternative questions for subsequent research: would the assessment change without the busiest week, the most active instrument, or one unusually large observation? Treat these as proposed diagnostic questions, not explanations already established by the data.

Repeat the review with a walk-forward schedule

In our proposed walk-forward procedure, each cycle has an earlier development segment followed by a reserved evaluation segment. Select the rule using only the designated development material, freeze its version, and then review the next segment. Repeat according to a schedule chosen in advance.

For a concrete planning example, develop on January through June and evaluate July. In the following cycle, either develop on February through July or retain January and extend through July; evaluate August. We call these our fixed-window and expanding-window conventions. Neither is presented here as the superior design.

Keep a row for every cycle, including the selected version, number of candidate configurations, evaluation dates, and decision. Preserve disappointing cycles in the report. If someone inspects an evaluation segment before freezing the relevant rules, our protocol withdraws that segment's reserved status.

At the end, write a narrative of the sequence rather than selecting only the most attractive cycle. Require separate notes on parameter changes and on changes to the research procedure. Under our convention, a changed research procedure starts a new documented series.

Illustrative schedule

Two cycles with reserved evaluation months

  1. January–June

    Develop the first candidate using the assigned six months.

  2. Before July review

    Freeze the first candidate and record its selection rationale.

  3. July evaluation

    Assess the reserved month and retain the complete report.

  4. Next cycle

    Develop on February–July, freeze again, and evaluate August.

Our planning example uses a fixed development window. The dates are assumptions, not a validated testing duration. First develop on January through June and evaluate July; then develop on February through July and evaluate August.

Paper trade with an explicit execution model

For this article, paper trading means recording hypothetical actions without submitting real orders. Our proposed journal can be maintained separately; we make no assumption that a particular MaxData simulation mode is available.

For each observed event, record the data timestamp, condition detection time, notification receipt time, and hypothetical execution time. Require an intended price, assumed execution price, fee allowance, slippage allowance, and reason for any skipped action. Mark every unmeasured item as an assumption rather than entering an unexplained zero.

The Polish MaxData backtest guide lists slippage, fees, and a spot-versus-perpetual market mismatch among possible reasons for differences between a backtest and live results. Execution limitations. Our proposed response is to maintain two explicitly labelled cost scenarios and to document assumed order size and liquidity conditions for each.

Include a reconciliation question: can another reviewer follow the recorded rule from the event to the hypothetical action? Require explanations for duplicates, missing notifications, and actions recorded after their stated expiry. Do not describe the journal as realistic unless you have first written a criterion for that assessment.

Use a written gate before any pilot

Our proposed gate has three outcomes: keep observing, permit a limited pilot, or reject the version. We do not prescribe a universal return target, minimum trade count, or acceptable drawdown. Choose organisational limits before reviewing the evidence, and describe them as your own requirements.

The proposed gate requires a complete experiment card, a frozen rule specification, a reserved-period report, the full walk-forward record, and a reconciled paper journal. Give each requirement an answer, a supporting document, and a named reviewer. An unexplained gap should remain visible in the decision record.

For a proposed pilot, document the allowed instruments, permitted actions, activity limits, review schedule, and person responsible for stopping it. Specify separately what should happen to actions already in progress when new actions are paused. Rehearse that decision as a written scenario or in an available test environment.

Under our checklist, a favourable historical report does not override an incomplete operational review. Record the decision date, approved version, and expiry of the approval. Any expansion in scope should receive a new review under this proposed process.

Our proposed decision convention

Choose and document one status

Keep observing
Decision reason
Required evidence remains incomplete
Next action
Record missing evidence and a review date
Limited pilot
Decision reason
All prewritten requirements are met
Next action
Record scope, limits, owner, and approval expiry
Reject version
Decision reason
The version fails a prewritten requirement
Next action
Archive the reason and identify any new research question
These are editorial review outcomes, not forecasts of performance. Three proposed review outcomes compared by decision reason and next administrative action.

Observe new events and define when the conclusion expires

Our live-observation stage means comparing newly observed events with the frozen specification. In this workflow, the label does not itself authorise order submission. Set a review rhythm, such as a daily journal check and a weekly version review, as an explicit scheduling choice.

Our proposed pause conditions include an unknown data source, delay beyond your documented limit, an unresolved execution state, or an action that cannot be reconstructed from the log. Proposed reasons to invalidate the research conclusion include changing rules during evaluation, excluding events without a recorded policy, or using a different market from the experiment card.

After a pause, require a written explanation and a decision about whether to repeat an earlier stage. Keep historical measurements and future expectations in separate fields. Allow “not specified” in the expectations field. End each review with the same questions: which data, which interval, what measured delay, what observation horizon, what limitations, and what would invalidate the conclusion?

Help resources and commercial invitation

The MaxData Help Hub presents an assistant for questions about MaxData and the workspace. Help Hub. Our suggested support question is: “Which document defines the input source and settings used in this report?” Ask for the underlying description so it can be attached to the experiment card.

Commercial invitation: consider creating a MaxData account. Our proposed research and approval decisions should remain documented in their own evidence file.

Sources

  1. MaxData Help Hub – Pomoc i przewodniki Mój workspace — agent.maxdata.app; accessed November 13, 2026
  2. Pro Alerts — agent.maxdata.app; accessed November 13, 2026
  3. Backtests — agent.maxdata.app; accessed November 13, 2026
  4. Backtesty Automatyczne Weryfikowanie Strategii — agent.maxdata.app; accessed November 13, 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
Backtest to live observation: an automation review workflow