Financial AI / AI Agent Research and Development

Why is a late re-entry dangerous — the background for designing GOLD's same-direction “up to 4 bars” rule

Even if you correctly determine the trend, if the time to re-enter in the same direction is late, the risk of reversal outweighs the remaining price range. In the development on September 17, we added a cap to GOLD13’s new-entry evaluation: only evaluate the first to fourth bars of the same direction; stop at entry from the fifth bar onward.

This article is a development log of indicators and automated trading software (EA) I developed, researched and validated with an AI agent and a real account. Not only successes, but also deferments, blocks, and failures are recorded with the same criteria. Validation status is also disclosed on a 24/7 weekday trading live stream.

The issue was not “direction” but “which bar number”

Even if the up or down direction is correct, in the late stage of a trend the room for growth is small, and a slight counter-movement can break the risk-reward balance. Traditionally, signals were judged mainly by signal strength and market direction, but the time axis of how many bars into the episode had been a missing piece.

Therefore, only candidates with the same symbol, same direction, fresh, REALTIME, warn_level=0 are counted. BUY and SELL are treated separately, and warn1/2, historical data, stale, and opposite direction are not counted. Even if there are duplicate records on the same bar, only one is counted. If the candidate interval exceeds 6 bars, it resets as a new episode.

The entry rules confirmed this time

• The first through fourth bars of the same episode proceed to new-entry evaluation
• From the fifth bar onward, blocked by both SOUL1 and NORMAL paths
• Save the blocking reason as guard_rejected / skip_warn0_late_cluster
• After restart, do not re-evaluate the same candidates
• After blocking, do not proceed to Gemini judgment, signal generation, or order submission

What matters is not only telling AI not to enter, but also persisting the fact and reason for the entry stop so that it can be audited later. This allows not only to evaluate favorable winning trades but also avoided loss candidates.

What was confirmed in the September 17 replay

The first bar of episode corresponding to ticket 29721052 was allowed, while the eighth bar corresponding to ticket 29721491 was blocked. The behavior including 4th bar allowed, 5th bar blocked, 6-bar reset, excluding warn1/2, history, stale, independent BUY/SELL, and non-duplicate on the same bar passed on 7/7.

The 21:15 scheduled run ended with no_breakout / exit=0, and there were no exceptions after loading the new source. However, passing in a replay and reproducing in live trading are separate proofs.

Things not yet proven

In live natural trading, it has not yet been confirmed that a fifth bar or later in the same direction with warn0 occurs and stops by skip_warn0_late_cluster. Therefore, separate “implementation complete” from “proved in a real account,” wait for the corresponding episode to occur, then record, telemetry, and order logs for cross-checking.

Evaluation axes readers should look at

The new filter cannot be evaluated simply by reducing the number of trades. We will measure how far the stop candidates revert or proceed in the same direction, the profit-loss difference with the first to fourth bars allowed, the rate of false stops, and repeated judgments after restarts. We will look at not only win rate but whether late entries limited the maximum drawdown.

There is no stop-motion video of corresponding natural trades yet, so unrelated recordings will not be embedded. When evidence is obtained, we will publish by aligning judgment times with charts.

GOLD13 Sales Page
Weekdays 24-hour trading live

This article is a research and validation log and does not guarantee profits.

Defining an episode to count the “5th bar”

Simply counting the number of entries for the day would treat morning rises and evening rises as the same flow. Therefore, an episode is defined as a contiguous interval of warn0 candidates in the same direction, switching to a new flow with a gap of at least 6 bars. If this boundary is ambiguous, normal initial movement may be misread as “late.”

An episode is not the truth of the market but a working hypothesis for validation. If real data shows that a 6-bar boundary is too short or too long, we will change it. The important thing is that the same data can be used later to reproduce the same count.

No interruption to existing position management even when the entry is stopped

Blocking new entries and managing SL/settlement for positions is separate. Even if the fifth bar onward is stopped, monitoring of existing positions, protection of profits, and limiting losses continue. Changes to entry rules are kept from propagating to exit rules.

Without this separation, a serious accident could occur where stopping new entries also stops performance management for existing positions. For an unsettled order at deployment, no policy changes or manual settlements will be performed, and conventional management will continue as the target.

Side effects of the episode limit

In strong trends, the fifth bar onward may still extend. The cap reduces losses but may cause lost profit opportunities. Therefore, we do not measure success solely by the number of blocks stopped; we track price movement after the stop.

• If the maximum forward movement after the stop is large, there is opportunity cost
• If the maximum backward movement is large, it helps avoid losses
• Liquidity factors when spread widens simultaneously
• If concentration occurs just before reversal, it is effective for late-stage detection

Roles of the AI agent and deterministic guard

Market environment classification and candidate prioritization are handled by the AI agent, while the guard ensures that counts and stops always yield the same results. We do not mix AI flexibility with EA reproducibility; prioritize explainability right before order placement.

Even if the model claims in words that “this time is an exception,” if the cap is exceeded, we will not proceed to order. If exceptions are allowed, we first modify the rules and validation items to avoid post hoc convenient interpretations.

Conditions for validation completion

In natural trading, fifth bar or later occurs, and candidate, guard, signal, order, and recording align in a single time series. Furthermore, gather multiple episodes to compare loss avoidance and opportunity loss at the stop. The two-stage condition must be met before labeling as “effectiveness verified.”

Verification table to be updated after publication

For each episode, summarize start time, direction, number of candidates, allowed/blocked, candidate price, spread, ATR, maximum forward/backward movement one hour later, and whether an actual order occurred, in one line. Not only successful blocking cases but also cases where the price extended after stopping will remain in the same table.

Until at least 10 episodes are collected, do not fix conclusions; compare the sensitivity of the 4-bar cap and the 6-bar reset. Do not overemphasize win rate with few cases; evaluate loss avoidance alongside opportunity loss. If the rules are changed, separate before and after the change date and do not mix past data with the new rule’s results.

Additionally, episodes with heartbeat gaps, recording gaps, or telemetry mismatches are deemed “undetermined” and will not be arbitrarily excluded. Not only the trading results but also the proportion of evidence that remains until the end will be published as a system quality measure.

Judgment criteria:Even after the article is published, as corresponding episodes increase, we will append in the same format, and for fewer than 10 cases we will indicate provisional results. The cap change will not be retroactively aligned with favorable trades, and the evaluation period and comparison items will be fixed before the change.