Regime-conditioned analogue forecasting of market risk, with self-exciting tail-risk gating. (StochQuant v2.)
A Rust + Python rebuild of the Stock_pine TradingView strategy ("Dynamic Stochastic Quantitative Trading Model Using 1D K-Nearest Neighbors and Poisson Risk Gating"), re-evaluated under a strict out-of-sample protocol, and redirected from direction forecasting to risk forecasting, which is where the data shows predictability.
The architecture of the original is kept: analogue (k-nearest-neighbour) matching, a Markov volatility regime, a Poisson-family tail-risk gate, and Kelly-style sizing. Each part was made statistically sound, then tested against standard benchmarks.
from analogue_risk import Analogue, volatility_forecast
from analogue_risk.data import binance
prices = binance("BTCUSDT", "1h") # open/high/low/close DataFrame
vf = volatility_forecast(prices, horizon=24) # next-day variance from volatility analogues
print(vf.tail())
# the general engine: any causal features, any targets
model = Analogue(horizon=5, k=20)
F = model.forecast(embedding, targets, state=regime) # mean/se per target, analogues used
idx, w = model.neighbours(embedding, targets, state=regime) # the analogues behind each forecast- API reference
- Tutorials (executed notebooks): 1. Forecasting volatility with analogues · 2. The method of analogues on the Lorenz system · 3. One-day VaR with analogue-conditioned FHS
The study was pre-registered: after a development study on 9 assets, eight hypotheses, the
models, a list of 15 new assets and the analysis code were committed and tagged
prereg-v1 before any of those assets' data was downloaded.
All eight replicated (results/confirmatory/hypotheses.md;
no deviations: DEVIATIONS.md).
| Question | Development (9 assets) | Confirmatory (15 new assets) |
|---|---|---|
| Does analogue matching predict direction? (H1) | No: mean IC −0.004, best deflated Sharpe 0.58 | No: mean IC +0.003 (p = 0.60), best deflated Sharpe 0.46 |
| Analogue+HAR vs HAR for volatility (H2) | tie (median QLIKE ratio 1.012); in 90% MCS 14/15 | tie (1.001); in MCS 23/25 (HAR 24, GARCH 20) |
| Analogue alone vs HAR (H3) | worse (1.103) | worse (1.087, Holm p < 0.001) |
| Gaussian kernel vs kNN analogue (H4) | better (0.967), found post hoc | better (0.953, Holm p < 10⁻⁶) |
| Gradient-boosted trees vs HAR (H5) | worse (1.52), in MCS 0/15 | worse (1.42), in MCS 2/25 |
| Hawkes vs the v1 static-Poisson crash gate (H6) | better (log-loss ratio 0.93) | better (0.85); v1 Poisson is worse than a constant in 20/25; a logistic on EWMA volatility beats both |
| Analogue/FHS blend vs FHS for VaR/ES (H7) | no difference | no difference; analogue VaR is as well calibrated as FHS, GARCH-t/FHS most accurate |
| Volatility targeting vs buy & hold (H8) | drawdown lower on 9/9 | drawdown lower on 15/15 (median −5.6 pp); Sharpe slightly lower (median −0.015) |
Also: the original "low-pass filter" (timeframe) claim gets only weak, non-significant support
(e6_timeframe); the analogue engine runs at ~30k bars/s with capped history and 200k bars in
~90 s with the full O(n²) search on 10 cores (e7_speed).
Package name. The Python package is
analogue_risk(it wassqup to the pre-registration;PREREGISTRATION.mdandresults/confirmatory/DEVIATIONS.mdkeep the original commands, e.g.python -m sq.experiments, which today ispython -m analogue_risk.experiments).
Analogue engine (src/analog.rs). For every bar, a causal embedding is compared with all past
anchors whose targets are fully observed. The k nearest per volatility regime are chosen greedily
under an exclusion zone (no two analogues within h bars, so their outcome windows do not overlap).
Regime groups are mixed by the h-step Markov forecast of the regime. A config guard rejects
settings where the exclusion zone would force non-neighbours into the set.
- Direction model: embedding = volatility-scaled log returns in m segments (m = 1 is the v1 feature); target = the R-multiple of the exact trade the strategy would take (same simulator as the backtester, so labels and execution cannot disagree).
- Volatility model: embedding = log variance of m past segments relative to the long-run level; target = log mean squared return over the next h bars (h = 24 hourly, 5 and 22 daily).
Benchmarks: EWMA (half-life selected per fold), GARCH(1,1) by MLE with variance targeting, HAR (Corsi 2009, log form), gradient-boosted trees (fixed hyperparameters), static Poisson (v1), Hawkes (exponential kernel, MLE), logistic on EWMA variance. For one-day VaR/ES: historical simulation, GARCH-N, GARCH-t, filtered historical simulation and analogue-conditioned FHS.
Protocol: first 30% of each series = calibration, the rest = 5 walk-forward folds. Every fitted
quantity uses only data observed before the fold. Model design was chosen on a validation slice
inside the calibration period (results/dev/design_study.*). Note for the paper: earlier pilot
runs looked at OOS numbers for BTC and the S&P 500 before the protocol was frozen; the frozen design
(kNN, k = 20, regime mixture, Analogue+HAR combination) was selected on calibration data.
Statistics: QLIKE and log-MSE; Diebold–Mariano with Newey–West (lag h); Hansen–Lunde–Nason Model Confidence Set (T_max, stationary bootstrap, 1000 draws); log-loss / Brier / AUC for tail probabilities; deflated Sharpe ratio for the direction grid; paired stationary-bootstrap tests of Sharpe differences.
Execution (direction backtests): signal at close, fill at next open, stops/targets intrabar with stop-first on ambiguous bars, gaps fill at the open, 5 bp per side (1 bp shown for v1 as in the original). Overlay: long-only, no leverage, 5 bp per unit turnover, 0.1 rebalance band.
The main study uses Yahoo's unadjusted prices. For the four distribution-paying funds (QQQ, TLT,
IWM, EEM; Yahoo's price indices have no adjusted series)
python -m analogue_risk.experiments --universe adjusted_etfs --only e2 e5 reruns the volatility and overlay
experiments on dividend-adjusted prices (results/adjusted_etfs/comparison.md).
Volatility results barely move (QLIKE ratios change by at most 0.026). Sharpe ratio levels rise
for both buy-and-hold and volatility targeting, most for Treasuries (+0.20, coupons), but every
comparison keeps its sign: volatility targeting still cuts drawdowns, lowers Sharpe slightly for
TLT, IWM and EEM, and raises it for QQQ.
The model confidence sets and Sharpe tests use fixed bootstrap block lengths (2h; 10 days for
VaR). --auto-block replaces them with Politis-White (2004) estimates
(results/block_length_comparison.md). No headline claim
moves: Analogue+HAR stays in the 90% MCS for 14/15 development and 23/25 confirmatory series, HAR
for 14/15 and 24/25. The automatic lengths are slightly more conservative, admitting a few more
models to some confidence sets (confirmatory: analogue alone 12 -> 14, GARCH 20 -> 22); 13 of 432
VaR/ES MCS verdicts and 3 of 168 overlay Sharpe-test verdicts change.
Following Chassot & Audrino (2026, IJF), --only e9 re-estimates HAR at every bar on rolling
windows fixed in advance (hourly 2190/8760 bars, daily 250/1000), alone and averaged with the
analogue (results/confirmatory/e9_rolling_har.md). On
this study's squared-return target, rolling HAR does not beat the expanding per-fold fit, and
the Analogue+HAR tie survives: against long-window rolling HAR the combination is statistically
indistinguishable (Wilcoxon p = 0.45 development, 0.47 confirmatory); against short-window rolling
HAR it is significantly better (confirmatory p < 1e-4), because averaging steadies a noisy fit.
The main study measures volatility with squared bar returns. For the eight crypto series,
--only e10 replaces them with realised variance from Binance 5-minute returns, as target and
as HAR/analogue input (results/confirmatory/e10_intraday_rv.md).
The tie holds: Analogue+HAR stays in the 90% MCS for 8 of 8 series and is never significantly
worse than HAR at 5% (BTC and ETH p = 0.08), and never significantly better (its loss is
slightly lower for three series and slightly higher for five). The analogue
alone is significantly worse than HAR in 6 of 8. With this cleaner target, rolling-window HAR
roughly matches the expanding fit, in line with Chassot & Audrino (2026). Caveat: 5-minute
realised variance exceeds squared hourly returns by 7-17% for liquid coins and 56% for the thinly
traded PAXG (microstructure noise).
The paper's results use the pre-registered kNN analogue (vol.HOURLY, vol.DAILY_W,
vol.DAILY_M), which stay unchanged so every number reproduces. The Gaussian kernel won the
pre-registered comparison with kNN (H4), so new work should start from vol.RECOMMENDED
(bandwidth=0.6, kernel_block=True). kernel_block treats each block of h consecutive anchors
as one observation when computing the effective sample size, because their outcome windows overlap;
it changes standard errors and the minimum-sample gate, not the forecast mean.
Binance access. Binance's REST API refuses some regions (e.g. HTTP 451 in the US). The loader
then falls back to the public bulk archive at data.binance.vision (or set
SQ_BINANCE_SOURCE=archive). The two official sources agree on 99.94% of bars but not all (the
archive contains a misaligned window during a February 2018 outage and a few bars differing around
exchange maintenance), so an archive-built series fails the snapshot check below and results can
differ marginally from the paper; the loader warns when this happens.
Pinned snapshot. python/analogue_risk/data_manifest.json records, for every series in the paper, the
last bar used, the row count and a SHA-256 of the cleaned data. Loaders truncate to that bar and
verify the hash, so re-running the experiments later reproduces the published numbers; if a
source has revised its history, a warning says so.
Binance spot klines (BTC, ETH, PAXG as a gold proxy; hourly, plus BTC 15-minute) and Yahoo daily
bars (S&P 500 from 1927, Nasdaq-100 ETF, GLD, TLT, EUR/USD, NIFTY 50), cached in data/. Yahoo
bars are passed through a documented bad-tick filter (|r| > 8 × 20-day σ and reversed > 70% on
the next bar); it removes three EUR/USD prints from 2008 and one S&P 500 print from 1935, and keeps
genuine events such as 2020-03-12.
pip install analogue-risk # once published to PyPI; prebuilt wheels for Linux, macOS, Windows
# or from a GitHub release: pip install analogue_risk-2.0.0-<platform>.whl
# or from source (needs Rust): pip install git+https://github.com/ManasDasri/analogue-riskAn installed package caches downloaded data in ~/.cache/analogue-risk (override with
ANALOGUE_RISK_DATA); a source checkout uses data/.
git clone https://github.com/ManasDasri/analogue-risk && cd analogue-risk
curl https://sh.rustup.rs -sSf | sh # Rust toolchain
uv venv --python 3.12 && uv pip install maturin numpy pandas scipy matplotlib tabulate pytest
.venv/bin/maturin develop --release # builds the Rust core into analogue_risk._core
.venv/bin/python -m pytest tests # 15 correctness tests
.venv/bin/python -m analogue_risk.experiments # development tables -> results/ (~35 min on an M4)
.venv/bin/python -m analogue_risk.experiments --universe confirmatory --only e1 e2 e3 e4 e5 e8
.venv/bin/python -m analogue_risk.hypotheses results/confirmatory # pre-registered tests
.venv/bin/python -m analogue_risk.figures # figures -> results/figures/
.venv/bin/python paper/make_tables.py # LaTeX tables for the manuscripts
cd paper/research && tectonic main.tex # or any LaTeX engine; also paper/softwarexTests cover: no look-ahead (forecasts, labels and the v1 port are identical on truncated data), Hawkes and GARCH likelihoods against brute force, Hawkes parameter recovery, label = executed trade, planted-signal recovery vs. random walk, MCS behaviour, and the bootstrap.
The Rust port of the original strategy (src/legacy.rs) can be checked against TradingView's own
backtest. In TradingView, add codes/script.pine (default inputs) to BINANCE:BTCUSDT, 1h, set
the chart timezone to UTC, open Strategy Tester -> List of Trades, export the CSV, save it as
tests/data/tradingview_btcusdt_1h.csv, and run
.venv/bin/python -m analogue_risk.pine_parity tests/data/tradingview_btcusdt_1h.csv # report
.venv/bin/python -m pytest tests -k tradingview # >= 95% of trades must matchsrc/ Rust core: analog.rs (engine), backtest.rs (simulator), hawkes.rs, garch.rs,
features.rs, legacy.rs (line-by-line port of the Pine v1 strategy), lib.rs (PyO3)
python/analogue_risk/ data.py, model.py (direction + v1 + gates), vol.py (risk models, overlay),
var.py (VaR/ES), stats.py, experiments.py, hypotheses.py, figures.py
paper/ research/ (journal manuscript), softwarex/ (software paper), refs.bib,
make_tables.py (every table generated from results/*.csv)
PREREGISTRATION.md frozen confirmatory plan (tag prereg-v1)
tests/ test_core.py
results/ development e1..e8 tables (.md + .csv), confirmatory/, figures/, config.json,
dev/ (calibration-period design study)
- Direction results are for price-only analogue features at 20-bar horizons; they do not rule out predictability from other information.
- Daily S&P 500 data before 1962 has close prices only (ATR regime uses close-to-close ranges).
- The overlay ignores financing and cash yield (no leverage, cash earns zero).
- The Analogue+HAR combination uses fixed equal weights; learned weights were not tested.
- PAXG is a gold-backed token, not spot XAU/USD; early PAXG history is thinly traded.
If you use this code, please cite the accompanying papers (in preparation) by M. G. Dasari, V. KVS and K. N. Meera, Amrita Vishwa Vidyapeetham, Bengaluru.
MIT © 2026 Manas Ganesh Dasari. The MIT license applies to the source code in this repository; the underlying method is also the subject of a patent by the authors.