Financial AI・AI Agent Research & Development / Weekly recap for the week of September 27, 2026 to October 3, 2026

Not Just Win Rate Left Behind | Reading Costs and Net Profit Separately in GOLD Real Account Verification

What I want to review first in this week's recap is not the number of wins, but what remains after deducting costs. GOLD trading involves large price fluctuations, and when profits concentrate in a short period, transaction counts and win rate can look favorable. However, in real-account verification, it is essential to keep gross, commission, swap, and net separately and show which numbers come from which primary data source; otherwise, the study's evaluation may be biased. This article does not aim to create new trading results. Using weekly figures re-aggregated from existing daily data as materials, it organizes how to read costs and unresolved items.

We are using Financial AI and AI agents for research and development of our own indicator GOLD13 and automated trading software (EA), and continue verification on real accounts. We also conduct 24/7 weekday live trading drive sessions to share progress as part of the research. However, the intraday unrealized P/L on the live screen, indicator displays, and EA trial logs are not identical to the post-settlement net profit. In the weekend summary, we prioritize the definition of records over the impressions from video or displays.

Scope covered in this weekly material

The editorial team’s primary data prepared at 18:33 JST on October 3 re-aggregates daily records from September 27 to October 2, recording 47 settled cases, 31 wins and 16 losses, gross 2967, commission -915, swap -289, net +1763. Arithmetic confirms that 2967 minus 915 and 289 equals 1763. The consistency between numbers is verifiable, but this alone cannot declare it as “the account’s realized profit,” because the primary data source clearly states currency units are not unified and broker statements from the same period have not been reconciled yet. It is necessary not to fill in currencies like yen or dollars or to reinterpret as the increase in account balance.

The 47 cases are a re-aggregation covering September 27 to October 2; the October 3 on-the-day close is not included. As of the data preparation, the final ledger, positions, results, and heartbeat for October 3 have not been verified. Even if the heading of the period includes October 3, the dates on which the results were confirmed are distinguished as up to October 2. In weekly articles, keeping the divergence between “aggregation period” and “target week” in the text helps track updates when numbers are added later.

Don’t hide the conversion from gross to net

Gross is the profit/loss calculated from the trades themselves; commission is the trading fee; swap is the adjustments due to rollover and similar factors. In this material, the sum of commission and swap is -1204, which is about 40.6% of gross 2967. This is the ratio obtained by dividing the figures listed in the material, and it does not mean the same burden applies to all accounts or all future trades. If you ignore costs and only look at gross, you will miss the distance to net +1763. At the same time, it is premature to conclude the superiority of a trading logic by focusing only on costs. Without aligning each ticket’s execution time, holding time, lot size, and fee structure, the cause cannot be identified.

In GOLD in particular, placing spread, slippage at order, commission, and swap into the same box makes it difficult to know what to improve. The spread appears as the difference between BID and ASK at market observation time, but the actual price difference borne depends on order direction and execution price. Commission and swap should be confirmed as separate items in the trading history. To avoid double counting costs absorbed at the execution price itself and costs appearing as a line item in the details, EA’s recorded items should also be separated. If numbers do not match, the AI agent should not automatically fill in convenient values, but pause for reconciliation as a pending item.

How to interpret 31 wins and 16 losses

If it’s just the count, it is 31 wins and 16 losses. But win rate alone does not reveal the average amount per winning trade, the average amount per losing trade, costs, the maximum drawdown, simultaneous holdings, or changes in lot sizes. In this material, among the 29 cases on October 1, 0.02 lots occurred in 14 cases with net +3293, and 0.04 lots occurred in 15 cases with net +7736. At first glance, larger lots seem better, but these tests were not conducted under the same time frame, same signals, or same stop-loss parameters. It is not appropriate to conclude to “increase to 0.04” when the trading conditions differ across the aggregated set.

On October 2, all 13 cases used 0.02 lots, with 8 wins and 5 losses, net -4603 in the material. Small lots do not inherently avoid losses, and even on days with gains, net profit can be negative. However, the ticket-level close reasons for which trades were settled are not ascertainable from weekly aggregation alone. Before reading the distribution of wins and losses, we check whether the same rules apply to the population or whether there were any sudden parameter changes.

Are you relying on one big daily gain?

The daily net on the daily data is -645 on September 28, -1063 on 29, -2955 on 30, +11029 on October 1, and -4603 on October 2. September 27 had zero newly settled closes. The total is +1763, but excluding the +11029 on October 1 leaves four days in the red. Describing this as “consistently winning over the week” would be inappropriate. Instead, the research task is to treat the fact that results depended on a large positive day and that a -4603 retracement occurred the next day as a matter to be studied. Rather than negating October 1’s result, we distinguish that reproducibility has not yet been demonstrated.

When looking at daily net fluctuations, it is necessary to consider not only the monetary amount but also how many trades settled that day, which lots were used, and which risk limits were effective. Simply comparing October 1, which mixed 0.02 and 0.04, with October 2, which had only 0.02, can make differences in market conditions, number of trades, and holding times appear as the effect of lot size. Financial AI should be tasked with documenting condition differences and withholding comparison rather than turning incomparable data into a beautiful graph.

Prioritize reconciling with broker statements

Next, it is necessary to uniquely map the 47 cases in the weekly material to the broker’s records for the same period. Compare in the same row ticket, symbol, trade direction, open/close times, lot size, execution price, gross, commission, swap, and net. Do not confuse JST with server time; preserve conversion rules. If values differ from the broker statements, do not express them as “confirmed” until the discrepancy can be explained. Even if numbers match, for the 47 cases where there is no video or screenshot in Drive’s master copy to verify entry and exit, you cannot claim that the footage confirms both entry and exit.

For our verification of self-created indicators and EA, a function to record the correspondence from signal to order is also necessary. However, the mere presence of incremental trial logs or a SELL display does not confirm that an order was placed. There can be order-cancellation, broker rejection, or communication delays along the way. For an AI agent to later explain the profit and loss, there must be a chain from signal ID to EA receipt, order response, ticket, positions, results, and ledger. If any section is missing, cost analysis remains provisional.

What to record in next week’s cost audit

Next week we will record, for each order, both the planned risk amount and the actual net. If the predetermined stop-loss position before the trade is unknown, we will not describe the result as “within expected loss.” We will also record the observed spread time, the actual open price, and the timing of commission and swap recognition separately. This will enable later examination of whether the price conditions were bad, whether costs increased due to holding the position longer, or whether the logic itself was flawed.

The daily report will include separate sections for “count,” “wins/losses,” “gross,” “commission,” “swap,” “net,” “lot distribution,” and “unmatched ticket count.” Do not substitute blanks with zero. Zero indicates that no costs occurred, and unknown indicates that it has not yet been read. An automatic aggregation that does not distinguish these two can accelerate misunderstanding, even if it increases the Financial AI’s processing speed. The weekday 24/7 live trading screen will also distinguish between shown unrealized P/L and confirmed net it presents.

This week’s conclusion

In this primary data, we have re-aggregated 47 cases, 31 wins and 16 losses, gross 2967, total costs -1204, net +1763. This is a useful intermediate result for research, but it cannot be called the final real-account performance until currency units and broker statement reconciliation are completed. It also shows that a large positive gain concentrated on one day while another day shows losses. The lesson for this week is not to hide costs, not to force comparisons of incomparable data sets, and to leave unresolved items as unresolved.

Research and development of self-made indicators, EA, and AI agents do not guarantee profits; they are an effort to make judgments verifiable from decision to result. To allow readers to verify tool content,GoGo Jungle product pageis provided. FX and GOLD trading carry the risk of losing capital, and this article does not recommend any individual trades.