Live against the simulator: five weeks of real trades, matched one to one

1. The question

The first two notes measured what the scored headlines know about price. This one asks a different question: when the index is traded with real money, does the account do what the simulator says it should? Every backtest on this site is a claim about what a rule would have done; the only check on that claim is a period in which the rule was actually running, with real fills, real fees and real latency, and a simulator run over the same days.

The production configurations, the roster of coins and the rule behind each, were finalised on 29 August 2026, when the trading bot was brought exactly in line with the simulator and the roster was fixed. From that day the live bot and the simulator are the same machine, and the comparison is clean. The backtest used here runs to 2 October, so the matching covers five weeks: 34 live trades entered between 29 August and 2 October and closed by 3 October. Four positions were still open on 3 October and are not counted.

2. What is compared

The live side is the bot's own ledger: every position it opened on Binance spot, with entry price, exit price, exit reason and timestamps, read from its database. Returns are net of 0.1% fee per side, the taker fee the account pays. Positions are equal in size and small against the order books of these pairs: fills, fees and latency are real, market impact is not tested.¹Three further coins (ATOM, AVAX, NEAR) run in signal-only mode: the bot publishes their signals and simulates the trades but does not place orders. They are in the simulator's output and excluded from the live comparison.

The simulated side is the trailing-stop backtest of the production configurations, the same run that feeds the deals on the Alpha evidence page: 30-minute candles, the pending-entry step waiting for the post-signal dip to reverse, exits by opposite signal or trailing stop, 0.1% fee per side, no slippage model.

Matching. A live trade is paired with the simulated deal on the same coin whose entry time is closest, within 36 hours, each simulated deal used at most once. For each pair four differences are recorded: entry time, exit time, entry price and exit price, and from them the difference in net return. A live trade with no simulated entry within 36 hours, or a simulated deal with no live entry, is reported as unmatched and explained.

3. Five weeks, trade by trade

Table 1. The window. Sums of per-trade returns are for orientation: positions are equal in size, so the sum is the return on one position's worth of capital turned over 34 times.

Of the 32 pairs, 19 lie within 0.5 percentage points of each other and 25 within one point. The median difference in net return is 0.02 points; the mean is +0.51 points in favour of live, with a standard deviation of 2.2 points, which is a paired $t$ of 1.3: no evidence that live is systematically better or worse than the simulation. The win rates over the window are 59% and 62%.

Entries agree to within an hour in 20 of 32 pairs and within six hours in 28; the median absolute difference is 25 minutes, which is the resolution of the two schedules. Live entries were on average 0.14% cheaper than simulated ones. Exits agree to the minute in 24 of 32 pairs. The exit reason, signal or trailing stop, is the same in 31 of 32.

Table 2. The window by coin. With two to seven trades per coin the per-coin means are not estimates of anything; the table is here to show that no coin is doing something the simulator does not.

4. Where they differ, and why

Five pairs differ by more than 2.5 points. All five are timing, and all five are visible in the two previous figures.

Later exits, better prices (three pairs). OP on 4 September: the live bot sold 36 hours after the simulator, at a price 8.5% higher, for +12.3% against +4.1%. AAVE on 15 September: 12 hours later, +4.4 points. UNI on 14 September: the simulator was stopped out by the trailing stop while the live bot held and sold on the next signal, +6.6 points; this is the one pair with a different exit reason. In each case the live bot and the simulator read a bucket boundary differently, the live bot stayed in, and the market happened to reward it. With the signs reversed these would have been the three worst pairs.

Earlier entries, worse prices (two pairs). ETH on 15 September: the live bot entered 5 hours earlier at a price 2.5% higher, −2.5 points. OP on 22 September: 9 hours earlier, 2.9% higher, −3.1 points. The live pending engine checks a real price every 30 minutes; the simulator sees candle closes. When a dip reverses sharply inside a candle the live bot fills first and higher.

Taken together the five explain the whole gap between +1.71% and +1.20% on the matched pairs. Remove them and the two means are 0.1 points apart with 27 pairs. The mechanism is not slippage, which on entries worked slightly in live's favour, but the granularity of time: a 12-hour signal clock, a 30-minute entry clock, and a bot that runs five minutes after the candle.

5. The entries that never happened

The production rule does not buy on the signal. It waits for the price to dip after the signal and turn back up, checking every 30 minutes, and gives up after 24 to 36 hours or when the strategy produces a SELL. In the window the bot received 76 BUY signals and filled 37. Twenty-two expired because the next SELL arrived before the dip reversed; seventeen timed out. The filled entries were bought a median 1.6% below the signal price, 11 hours after the signal, after a dip whose low point was a median 2.6% below it.

The simulator runs the same rule, and this is where the two unmatched sets come from. Seven simulated deals on the traded coins have no live counterpart: three (ETH 30 Aug, LINK 19 Sep, AAVE 20 Sep) because the live pending entry was cancelled by a sell signal while the simulated one had already filled on a candle close, a difference of minutes at the boundary; two (UNI on 29 and 30 September) that fell while the live bot was being stopped out of the previous UNI position and its next entry waited for a dip that came on 2 October; and two (BTC and OP, both on 15 September) where the noon signal was marginal: the simulator opened on it and closed at the next bar for +0.5% and −0.5%, while the live bot, which computes the score five minutes after the bar on the headlines ingested by then, landed on the other side of the threshold. Together the two would have added −0.04 points to the window. Two live trades have no simulated counterpart (LINK 13 Sep, ETH 25 Sep), the mirror case. Of the 39 simulated entries on the traded coins, 32 happened live, 82%; of the 34 live entries, 32 are in the simulation, 94%.

6. What the five weeks earned

Over the window the live account closed 34 trades on the seven coins at an average of +1.73% net of fees, 59% of them profitable, +58.7% summed across trades of equal size; the simulation of the same configurations over the same days made +1.44% per trade on 39 trades, +56.3% summed. The last eight live trades, entered on 28 September and after, averaged +3.4%, with two AAVE trades at +14.6% and +14.4% and one UNI trailing stop at −4.6%; the simulator has both AAVE trades too, at +15.5% and +13.7%. Four positions were open on 3 October.

Table 3. The window, net of fees. The last week is shown because it is the most recent, not because it is representative: eight trades.

On the account as a whole, with positions of about an eighth of equity each and several open at a time, this is a return of about eight percent over the five weeks; the exchange's own spot PnL statement puts the account at +8.4% from 25 August to 3 October, the span that contains the window. The number is given for completeness and should be read with the sample size in mind: five weeks, one market regime, and a strong final week.

7. What this says

The simulator is faithful. In the first five weeks in which the two ran identical rules, 32 of 34 live trades have a simulated twin, entries agree to 25 minutes, exits to one minute, exit reasons in 31 of 32, and the median return difference is 0.02 points. Nothing in the live account is happening that the simulation does not predict. The remaining differences are the granularity of time, and they cut both ways.

The five weeks were positive, and that is five weeks. +1.7% per trade over 34 trades, about eight percent on the account, with the strongest week at the end, against +1.4% per trade in the simulation of the same days. That is consistent with what the backtest of these configurations says they earn, and it is far too short to confirm it. This note will be extended as the months accumulate, with the same matching and the same tables.

Limitations. Five weeks and seven coins. Matching within 36 hours is a choice; a tighter window changes nothing in the pairs but would move the two unmatched live trades. Fees are taker fees at the account's tier; slippage is measured, not modelled, and was negligible at this size. The configurations were chosen in-sample on the two years of backtest, and the live window overlaps the end of that sample. Every pair and every unmatched trade is in the data files with its timestamps.

8. Update, 25 October 2026: eight weeks, 62 trades

The note above was written on 3 October with five weeks of trades. This section repeats the comparison on the window from 29 August to 24 October, eight weeks, with the same matching rule and the same configurations, which have not changed. Nothing above has been edited.

Table 4. The eight-week window, net of fees. The simulator also runs three configurations the live account does not trade, ATOM, AVAX and NEAR; their 24 deals are in the last column and have no live counterpart by construction.

Fifty-seven of the 62 live trades have a simulated twin, 92%. On the matched pairs the live mean is +0.76% against +0.63% simulated, a difference of 0.13 points with a standard deviation of 2.2 points, paired $t = 0.5$; the median difference is $-0.02$ points. Thirty-five of the 57 pairs are within half a point of each other and 44 within one point. Entries agree to within an hour in 36 pairs and within six hours in 49; the median absolute difference is again 25 minutes, 26 live entries earlier than the simulator and 31 later. The live entry price is a median 0.06% above the simulated one, mean 0.09%, with 29 of 57 pairs within a quarter of a percent either way and the 10th and 90th percentiles at $-0.57\%$ and $+0.79\%$. Exits agree in reason in 55 of 57 pairs and in price to within half a basis point in the median.

Table 5. The eight-week window by coin. Signs and order agree with the simulator for every coin; the one gap in level is OP, where the single pair of 4 September described in Section 4 accounts for 8 of the 12 points of difference between the two sums.

The large differences are the ones Section 4 already named plus three new ones, and they are the same mechanism. SOL on 8 October: the live bot entered 31 hours before the simulator at a price 3.3% higher and exited 36 hours after it at a price 4.6% lower, $-7.8$ points, the worst pair in the sample. UNI on 2 October: the live bot was stopped out, the simulator held through the dip and sold on the signal 66 hours later, $-4.6$ points. UNI on 14 October: both stopped out, the live stop 3.1% higher, $+3.2$ points. Remove the seven pairs that differ by more than 3 points, the three above and four of the five from Section 4, and the two means are 0.01 points apart on the remaining 50, as before.

The entry engine over the eight weeks: 138 BUY signals, 69 filled, 69 expired, 42 of them because a SELL arrived first and 27 by timeout. The filled entries were bought a median 1.8% below the signal price, 11.6 hours after it, after a dip whose low point was a median 2.7% below the signal; the expired ones waited a median 24 hours. The 17 simulated deals on the traded coins without a live twin and the 5 live trades without a simulated one are the boundary cases of Section 5, mostly UNI and SOL.

Table 6. The live account by week of entry, net of fees. The strong week of 28 September that closed the original window was followed by two negative ones; over eight weeks the mean per trade is +0.79% against +1.73% over the first five.

What changed with three more weeks is the level, not the agreement. The live account earned less, +0.79% per trade instead of +1.73%, because October was a weaker month for the configurations, and the simulator on the same seven coins earned less too, +0.34% instead of +1.44%. The two still match trade for trade: 92% of live trades have a twin, the twins exit for the same reason 96% of the time, and the median price difference at entry is six basis points. That number, the realised cost of executing a simulated entry, is what the next note needs. This section will be extended as the record grows; the window, the matching rule and the configurations stay fixed.