MatchedFlowConformalSignal

MatchedFlowConformalSignal is an intraday OHLCV signal that tries to trade only when directional price movement, signed volume flow, VWAP location, and relative activity point in the same direction and the expected move is larger than a recent empirical error band.

It is a composite research prototype rather than a direct implementation of one paper. The main references are:

• Chordia, Roll, and Subrahmanyam, "Order imbalance, liquidity, and market returns," Journal of Financial Economics, 2002. The paper studies order imbalance as a trading-activity measure and links imbalance to liquidity and market returns. Public metadata: https://authors.library.caltech.edu/95132/
• Chordia and Subrahmanyam, "Order imbalance and individual stock returns: Theory and evidence," Journal of Financial Economics, 2004. The paper studies how autocorrelated imbalances can create price pressure in individual stocks. SSRN page: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=354122
• Xu and Xie, "Sequential Predictive Conformal Inference for Time Series," arXiv:2212.03463, 2022. The paper motivates sequential calibration of nonconformity scores for non-exchangeable time series. arXiv/HF page: https://huggingface.co/papers/2212.03463

The RTTA implementation is deliberately simpler than the papers. It does not estimate true order imbalance from signed trades or quotes. Instead, it uses the bar close-to-close return sign as a cheap proxy for signed flow and scales that by dollar volume.

API

rtta.MatchedFlowConformalSignal(
    horizon_bars=12,
    calibration_window=250,
    calibration_quantile=0.80,
    entry_z=1.0,
    cost_buffer=0.0005,
    max_abs_target_fraction=0.05,
    participation_cap=0.02,
    fillna=False,
)

update(open, high, low, close, volume, normal_dollar_volume=nan, market_cap=nan, reset_session=False) consumes one aggregate bar. The example script uses Massive/Polygon stock trade aggregates, but any regular OHLCV bar stream can be used.

State

The indicator maintains:

• close lags for 3, 6, and 12 bar momentum,
• rolling return volatility over 6 and 12 bars,
• rolling sums of alpha_flow over 3, 6, and 12 bars,
• rolling sums of participation flow over 3 and 6 bars,
• an intraday session VWAP built from bar VWAP or close times volume,
• an EWMA fallback for normal dollar volume,
• a fixed-size queue of pending predictions,
• a rolling quantile of realized prediction errors.

reset_session=True clears intraday state such as close lags, flow windows, pending predictions, and session VWAP. It does not clear the longer-lived normal-dollar-volume EWMA or the calibration residual quantile.

Feature Construction

For each valid bar:

1. Compute dollar_volume = close * max(volume, 0).
2. Choose normal_dollar_volume from the supplied value, the internal EWMA, or current dollar volume.
3. Compute close-to-close log return ret1.
4. Use sign(ret1) as the signed-flow direction.
5. Update session VWAP from cumulative price-volume and volume.
6. Compute: - vwap_gap = close / session_vwap - 1 - rel_dollar_volume = dollar_volume / normal_dollar_volume - alpha_flow = sign(ret1) * dollar_volume / market_cap, when market cap is supplied, otherwise scaled by normal dollar volume - participation = sign(ret1) * dollar_volume / normal_dollar_volume
7. Clamp alpha_flow and participation to [-10, 10].
8. Push these values into the rolling flow windows.

Momentum is a weighted combination of 3, 6, and 12 bar log returns:

momentum = 0.20 * ret_3 + 0.35 * ret_6 + 0.45 * ret_12

flow_score compresses accumulated signed flow through tanh, with a different normalization depending on whether market_cap is supplied:

flow_score = tanh(alpha_flow_12_sum / alpha_norm
                 + 0.50 * participation_6_sum / 6)

Relative activity is also squashed:

activity_score = tanh((rel_dollar_volume - 1) / 2)

The high-low range dampens predictions in wide-spread/noisy bars:

spread_proxy = max(0, high - low) / close
spread_dampen = 1 / (1 + 25 * spread_proxy)

Prediction

The raw next-horizon log-return forecast is:

raw_prediction =
    0.35   * momentum
  + 0.0010 * flow_score
  + 0.05   * vwap_gap
  + 0.0005 * activity_score

prediction = spread_dampen * raw_prediction

This is intentionally heuristic. The goal is not to fit a model in the indicator; the goal is to expose a stable, incremental signal that reacts only when several intraday conditions line up.

Error Band

Every prediction is stored with its entry close. After horizon_bars updates, the indicator compares the realized log return to the stored prediction:

realized = log(current_close / old_close)
error = abs(realized - old_prediction)

Errors are pushed into a rolling quantile. Once enough errors are available, radius becomes:

radius = max(rolling_error_quantile, cost_buffer)

Before calibration is ready, it uses a fallback:

radius = max(2 * cost_buffer,
             max(0.0007, 1.25 * volatility_12 * sqrt(horizon_bars)))

This is conformal-inspired but not a formal finite-sample conformal guarantee. It is an online empirical error band.

Trading Outputs

The score is:

score = prediction / (radius + cost_buffer)

The signal is:

signal = +1 if prediction > entry_z * (radius + cost_buffer)
signal = -1 if prediction < -entry_z * (radius + cost_buffer)
signal =  0 otherwise

The target fraction is capped:

target_fraction = max_abs_target_fraction * clamp(score / 3, -1, 1)

max_trade_dollars = participation_cap * normal_dollar_volume is a liquidity cap for an external execution layer.

Outputs

• prediction: expected next-horizon log return.
• radius: empirical uncertainty/error band.
• score: prediction divided by radius plus cost buffer.
• signal: -1, 0, or +1.
• target_fraction: capped suggested exposure.
• alpha_flow: signed dollar volume scaled by market cap or normal dollar volume.
• participation: signed dollar volume scaled by normal dollar volume.
• flow_score: compressed recent signed-flow measure.
• momentum: weighted multi-bar momentum.
• volatility: rolling 12-bar return volatility.
• vwap_gap: close relative to session VWAP.
• rel_dollar_volume: current dollar volume relative to normal dollar volume.
• max_trade_dollars: liquidity cap.
• realized_error: newest matured prediction error, when available.

Intended Use

Use this on liquid intraday bars, typically 1 to 5 minutes. Reset it at session boundaries. It works best as a ranking or gating feature inside a broader system that also handles spreads, fees, inventory, exposure, and execution.