Thales
← research journal
Nov 13, 2026raw markdown ↗

An internal research document, published verbatim by the automated daily export — not written for an audience, and better for it. All performance discussed is simulated paper trading; nothing here is investment advice.

2026-11-13: what the robo rule would have done, 2007 to today

This is for expectations, not evidence. The mix, the rule and the checks were pinned and pushed in commit de508cf before this was run, and no pinned value has moved since. A fixed mix has no edge to prove; the point of the history is to know what an ordinary year, a bad year and a harvest look like before the live record starts.

Reproduce: python research/robo_history.py --fetch. It runs simulate(), which calls the same decide() the live sleeve calls, on config/robo.yaml as committed.

Re-run once, after the code review of the same day. The first run used a rule whose bands were centred 0.3 points below the spec's table (on the target less the cash reserve, so 54.7% to 64.7% for US stocks). The review caught the mismatch and the code now matches the table (55% to 65%). The figures below are from the corrected rule. The first run's are kept in the last section so the difference is on the record.

The short version

  • It tracks its own mix closely. Over 18 full years the rule finished within 0.56 points of its yardstick every year (average gap +0.06). The pre-registered limit is 0.75.
  • It trailed SPY in 14 of those 18 years, by 3.0 points a year on average. The whole of that gap is the mix: international stocks and Treasury bills did worse than US stocks over this stretch. It is a fact about the last two decades, not a forecast.
  • It falls with the market. The worst fall was 52.7% (October 2007 to March 2009), against 55.2% for SPY. Nothing in the rule softens that, by design.
  • It barely trades. Fifteen rebalances in 19.5 years, moving about 5% of the account in a typical year.
  • Tax-loss harvesting pays early or not at all. All 19 harvests fell in 2007 to 2009, right after the simulated purchase. In 16 of 18 full years there were none, including the 2020 and 2022 falls, because by then the funds were far above what was paid for them.

The numbers

Window 2007-06-01 to 2026-11-13, one trade a day at the close, 1 basis point each way.

Per yearVolatilityWorst fall
Robo rule8.1%17.9%−52.7%
Its yardstick (60/30/10, rebalanced daily, no costs)8.0%18.0%−52.7%
SPY10.5%19.7%−55.2%
YearRoboYardstickGapSPYRobo − SPYRebalancesHarvests
2007 (from June)−1.4%−1.0%−0.35−3.7%+2.414
2008−35.3%−35.3%−0.05−36.8%+1.5212
2009+29.3%+28.9%+0.40+26.4%+2.923
2010+13.9%+14.2%−0.28+15.1%−1.110
2011−3.5%−3.7%+0.17+1.9%−5.400
2012+15.5%+15.5%−0.03+16.0%−0.510
2013+23.8%+24.0%−0.17+32.3%−8.510
2014+6.1%+5.9%+0.15+13.5%−7.400
2015−1.0%−0.9%−0.12+1.2%−2.300
2016+9.3%+9.2%+0.02+12.0%−2.710
2017+21.2%+20.9%+0.31+21.7%−0.510
2018−7.3%−7.3%−0.07−4.6%−2.800
2019+25.2%+24.9%+0.23+31.2%−6.100
2020+17.0%+16.4%+0.56+18.3%−1.330
2021+18.1%+17.9%+0.24+28.7%−10.610
2022−16.2%−16.3%+0.04−18.2%+2.010
2023+20.7%+20.9%−0.14+26.2%−5.500
2024+16.4%+16.2%+0.21+24.9%−8.510
2025+20.0%+20.4%−0.34+17.7%+2.300
2026 (to 11-13)+10.6%+10.6%+0.03+11.5%−0.900

The first rebalance in 2007 is the initial purchase.

Reading the pre-registered checks against this

CheckLimitIn this history
Return fidelitywithin 0.75 points a yearworst full year 0.56 (2020)
Rebalancing turnoverunder 50% a yearworst year 22.1% (2020, three rebalances), average 5.5%
Buys of a locked fundzerozero, across 19 harvests

None of the limits was set from these figures, and none is changed by them. The fidelity limit is neither loose nor tight here: a gap of 0.2 to 0.5 points is normal, because the live rule lets weights drift inside their bands while the yardstick is rebalanced every day.

The harvests

The simulated account bought in June 2007, four months before the peak, so it is close to a worst case for timing and a best case for harvesting. It swapped funds 19 times through the fall, about once a month per stock class while prices kept dropping, and banked losses worth about two-thirds of the account in total (measured against the account on each harvest day). At the pinned 24% rate that is an estimated tax value of 16% of the account, nearly all of it in 2008.

Three cautions on that figure:

  • It assumes every dollar of loss can be used. In practice a loss offsets gains first, then $3,000 a year of other income, and the rest carries forward.
  • It is a deferral. The replacement funds were bought low, so their eventual sale has a larger gain.
  • After 2009 the count is zero for 17 years. An account that starts in December 2026 harvests only if the market falls below where it bought. A flat or rising first year shows no harvests, and that is the rule working.

What this history cannot tell you

  • Whether the mix beats SPY from here. US stocks led for most of this window; the stretch before it went the other way. The mix was chosen so that the sleeve does not need to know.
  • Real harvest frequency. These prices fold dividends in, so funds look less far below cost than a real account's price-only cost would show. Real harvests would come somewhat more often.
  • Mid-morning execution. The simulation trades at the close; the live sleeve trades at the broker's quote around 10:35 ET.
  • The funds before they existed. VXUS (2011), IXUS (2012) and SGOV (2020) are extended backwards with VEU, CWI and BIL, which track the same markets.

The first run, for the record

Before the band correction the same script gave: 8.0% a year; yearly gap to the yardstick between −0.42 and +0.44 (average +0.05); worst turnover year 11.4%; the same 15 rebalances on slightly different dates; the same 19 harvests; the same worst fall; the same 14 of 18 years behind SPY. The correction moved rebalance dates by 0.3 points of drift and nothing else.