finmlsim 0.4.1

Simulation toolkit for financial machine learning: scholarly market, microstructure, and execution simulators for teaching and research.

finmlsim

Scholarly simulators for financial machine learning — a small, NumPy-only package of citation-anchored simulation models for research, teaching, and self-study. Twenty-plus generators across classical return series (Gaussian, Student-t, ARMA, GARCH, regime-switching, jump-diffusion, Cont–Bouchaud), continuous-time price processes (Heston, Bates, SABR, rough Bergomi, Ornstein–Uhlenbeck / Vasicek, fractional Brownian motion), and market microstructure / execution (Hawkes, Roll, Glosten–Milgrom, Kyle, a Cont–Stoikov limit-order book, Almgren–Chriss). Each model is one short function with a citation in its docstring.

Dial any stylized fact of markets on and off — fat tails, volatility clustering, jumps, regime switching, long memory, order-flow self-excitement, adverse selection — and watch what the methods you care about do under those conditions, before you turn them loose on real data.

import finmlsim as fms

# Volatility clustering plus fat tails
r = fms.simulate.garch(n=2000, dist="t", seed=0)
fms.stylized.summary(r)
prices = fms.stylized.to_prices(r)

# Stochastic-vol pricing with Heston
S, v = fms.simulate.heston(n=252, seed=0)

# Cont-Stoikov limit-order-book dynamics
book = fms.simulate.limit_order_book(n_events=20_000, seed=0)

Install

pip install finmlsim
# core needs only numpy; analysis extras: pip install "finmlsim[analysis]"

Modules

• simulate — generators grouped by family:
  • Return series (1-D log returns): gaussian, student_t, arma, garch, regime_switching, jump_diffusion, cont_bouchaud.
  • Cross-sectional panel: panel (2-D, common-factor structure).
  • Continuous-time price processes: ornstein_uhlenbeck / vasicek, heston (1993), bates (1996), sabr (Hagan et al. 2002), rbergomi (Bayer-Friz-Gatheral 2016), fbm / fgn (Hurst-parametrized).
  • Microstructure and execution: hawkes (1971 self-exciting), roll_bounce (1984), glosten_milgrom (1985), kyle (1985), limit_order_book (Cont-Stoikov style), almgren_chriss (2000 optimal execution). Every generator takes a seed for reproducibility.
• stylized — diagnostics: acf, excess_kurtosis, vol_clustering, summary, to_prices.
• metrics — strategy evaluation: sharpe, max_drawdown, rank_ic, ic_ir, turnover, deflated_sharpe, min_track_record_length.
• resample — block bootstraps for dependent series: stationary_bootstrap (Politis-Romano 1994), circular_block_bootstrap (Politis-Romano 1992). Pairs with the multiple-testing diagnostics in metrics.
• data — loaders for the bundled empirical datasets (Ken French daily factors, momentum, and 12 industry portfolios): ff_factors, ff_momentum, industries, market_return. Offline; see datasets/famafrench/ in the source repository.

The stylized facts each return-series generator produces

Generator	Fat tails	Vol clustering	Mean autocorrelation
gaussian	–	–	–
student_t	✓	–	–
arma	–	–	✓ (by design)
garch	mild	✓	–
garch(dist="t")	✓	✓	–
regime_switching	✓ (mixed)	✓	–
jump_diffusion	✓	–	–
cont_bouchaud	✓ (endogenous)	(weak)	–

What the continuous-time and microstructure models give you

Generator	Returns	Key parameter	Use case
ornstein_uhlenbeck / vasicek	path	kappa (mean-reversion rate)	pairs, short rates, vol state
heston	(S, v)	xi, rho, kappa	stochastic vol; smile/skew
bates	(S, v)	+ jump_intensity	stoch-vol with crash risk
sabr	(F, alpha)	beta, nu, rho	FX / rates vol surfaces
rbergomi	(S, xi)	H < 0.5	rough volatility
fbm / fgn	path / increments	H	Hurst-parameterized paths
hawkes	event times	alpha, beta	order-flow clustering, contagion
roll_bounce	tx-price log returns	half_spread	bid-ask bounce, negative AC(1)
glosten_milgrom	dict (ask/bid/spread/belief)	alpha (informed prob)	adverse-selection spread
kyle	dict (v, u, x, y, price)	sigma_v / sigma_u	strategic informed trading
limit_order_book	dict (mid, spread, depth)	lam0, theta, mu_mkt	LOB dynamics, depth, impact
almgren_chriss	dict (path, costs)	lam (risk aversion)	execution trajectories

These are teaching and research models, not production calibrations. Planned additions: a cost-aware backtest module and leak-free features helpers.

License

MIT. See LICENSE in the source repository.