Skip to the page
MODEL AND VALIDATION/DEFAULT MARKET MODEL

Default market model

In tradefloor 0.11.0, a market runs on the preset pt-v22 when you don't pass a model. pt-v22 passes 190 of the 190 validation criteria that count toward its grade, which were registered before grading. The grade ran 270 simulated histories of 21 years on holdout seeds the tuning never used. On the one-year table, all 19 statistics fall inside acceptance ranges read from real market data. pt-v21 and every earlier preset stay available by name. In the preset table at the end, 15/15 means all 15 fixed-roster statistics a record measured fell inside their ranges. How it is measured lists each criterion with its observed result.

What it gets right

Crash rates inside the registered band

Over 270 simulated histories of 21 years, the index has 1.69 bear markets of 20% a decade against a real 1.16, 4.62 corrections of 10% against 2.84, and 7.2 sessions a decade down more than 5% against 6.2, all within the half-to-twice band the rows allow. In replays of 2008 and 2020 with the real VIX imposed, the maximum drawdown is 44% and 34% against a real 57% and 34%.

How often and how long the VIX stays high

The VIX is above 30 on 6.8% of sessions against a real 8.2%, and a spell above 30 lasts 24 sessions on average against 22.

Rates and bonds that move prices

Driven through the real rate path from December 2021 to October 2022, the market P/E moves -2.08% per 100 bp of the corporate yield, against the S&P 500's -5.43%. The simulated 10-year Treasury moves 5.89 bp a day against a real 5.41, and the correlation of stock and bond returns has the sign measured in real data.

Earnings that fall in recessions

Around a contraction, the index's reported earnings move -9.1% against -17% in Shiller's data, and after the low of the packaged 2008 recession (recession_2008) the index rises 61% within a year, against 69% after March 2009.

Costs that grow with size

The cost of an order grows with the square root of its size. The fitted exponent is 0.46, against the empirical 0.5.

Leak checks

Some checks are strategies built to find leaks, because a market that's easy to game makes a bad strategy look good and teaches a learning agent an edge real markets don't have. Reading a headline 5 ticks late earns 17.9 bp, under the 20 bp limit. No price-only rule on the published suite of 20 markets beats buy-and-hold by more than 5 points. Value and momentum screens stay inside bands set from real data, and timing the market on published economic data earns at most a point a year.

Long-run return and volatility

The long-run index return is 7% a year against a target of 6.25%, with annual volatility of 18.4% against 18.1%.

What a simulated market adds

If you fork a market at day 100 and raise rates in one copy, the difference between the copies is the effect of the rate rise on your strategy, with luck held fixed. Running an agent on 30 seeds gives you 30 versions of the economy to test it on, each with its own crashes and recoveries. A result tells you how a strategy behaves in this model and doesn't forecast returns in a real market.

What changed from pt-v20

  • The business cycle feels the stock market. A fall below the index's slow average raises the chance of a downturn.
  • The Fed cuts into a falling market and into stress, and the Treasury curve prices the policy path it expects.
  • Credit spreads follow the cycle, the VIX and how leveraged companies are.
  • Companies pay dividends and buy back shares out of earnings, and a stock drops by its dividend at the open.
  • Market volatility follows the cycle, rises more after falls than after rises, and has a slow part that lasts for months.
  • Each run starts after two years of market history, so day 0 opens with the volatility and valuation a market would have.
  • Each company's own volatility clusters, and market jumps are rarer, larger and mostly down.
  • Part of each day's move happens overnight, between the close and the next open.
  • Companies report earnings on a calendar, with surprises, a move on the day and drift before and after.
  • The cost of trading remembers recent volume, so splitting a large order saves some of its cost.
  • Unemployment follows growth, and oil reverts to its inventory level and feeds into inflation.

pt-v20 and all earlier presets keep their values. model="pt-v20" still runs pt-v20 and replays it exactly.

Limits

The named gaps lists each place the model measurably falls short, with the uses it rules out. tf.envelope.check() refuses a question that depends on one.

Every preset

A result names the preset it ran on, and since shipped presets don't change, it replays on that preset in every later release. Pick one with model="pt-v20" and so on (the numbering skips 17). The counts in the table are each record's fixed-roster panel of shape statistics, at one and two years.

PresetStandingIn band, one yearIn band, two years
pt-v22Default from 0.11.015/1515/15
pt-v21Default from 0.10.0, replaced in 0.11.015/1515/15
pt-v20Default from 0.8.5, replaced in 0.10.015/1515/15
pt-v19Default from 0.8.0, replaced in 0.8.515/1515/15
pt-v18Default from 0.7.0, replaced in 0.8.014/1414/14
pt-v16Default from 0.6.0, replaced in 0.7.014/1414/14
pt-v15Never the default14/1414/14
pt-v14Default from 0.4.0, replaced in 0.6.014/1414/14
pt-v13Never the default14/1414/14
pt-v12Default from 0.3.0, replaced in 0.4.014/1413/14
pt-v11Never the default13/1412/14
pt-v10Default from 0.2.0, replaced in 0.3.013/1413/14
pt-v9Never the default13/1412/14
pt-v8Never the default13/1413/14
pt-v7Never the default13/1413/14
pt-v6Never the default12/1410/14
pt-v5Never the default12/149/14
pt-v4Never the default11/149/14
pt-v3Default from 0.1.0, replaced in 0.2.012/148/14
pt-v2Never the default11/149/14
pt-v1Never the default9/148/14