Standalone features that can be fed to an ML model.

Setup

These headers include the amalgamated single-header build of QModel, not the multi-file src/ version. You need to generate it once:

1. In QModel/, run ./amalgamate.sh. This produces a QModel/ folder containing QModel.h and the GLFW headers.
2. Copy (or symlink) that QModel/ folder into this directory as vendor/, so you end up with Features/vendor/QModel.h.

Including a feature

Just include the file you want and call the wrapper:

#include "priceReturnFeature.h"

AgentParameter returnHorizon(10.0, 1.0, 100.0, true);

std::vector<double> features = priceReturnFeature(state.idx, returnHorizon);

Multiple features combine by concatenating:

#include "priceReturnFeature.h"
#include "cumulativeVolumeDeltaFeature.h"
#include "bookImbalanceFeature.h"

std::vector<double> features = priceReturnFeature(state.idx, returnHorizon);
auto cvd = cumulativeVolumeDeltaFeature(state.idx, cvdHorizon);
auto book = bookImbalanceFeature(state.idx); // no horizon needed

features.insert(features.end(), cvd.begin(), cvd.end());
features.insert(features.end(), book.begin(), book.end());

Most features take an AgentParameter horizon so the model can tune the lookback window itself, same as derivativeHorizon in main.cpp. A couple (bookImbalanceFeature, tickRunLengthFeature uses it as a lookback cap rather than a fixed window) don't need one.

The features

priceReturnFeature - point return from horizon bars ago to now. Plain momentum: positive means price climbed, negative means it fell.

rollingVolatilityFeature - stdev of tick-to-tick point returns over horizon bars. A vol-regime read.

zScoreFeature - how many stdevs the current price sits from its rolling mean over horizon bars. Mean-reversion signal, big magnitude = stretched.

relativeStrengthFeature - RSI rescaled to 0-1 (0.5 = neutral) over horizon bars of gains vs losses.

cumulativeVolumeDeltaFeature - sum of orderflow delta (executedBuys - executedSells) over horizon ticks. Raw CVD, positive means buyers are in control.

vwapDeviationFeature - distance in points between the current price and the volume-weighted average price over horizon ticks.

bookImbalanceFeature - resting bid/ask size imbalance at a single tick, +1 stacked on the bid, -1 stacked on the ask. No horizon, it's a snapshot.

aggressorRatioFeature - net aggressor split over horizon ticks, +1 = everything traded on the ask (all buying), -1 = everything hit the bid.

volumeSpikeFeature - current tick's volume vs its rolling average over horizon ticks. 1.0 = normal, above 1 = a spike, below 1 = quiet.

tickRunLengthFeature - length of the streak of consecutive up-ticks or down-ticks ending at idx, signed by direction, capped at maxLookback.

deltaDivergenceFeature - flags price/orderflow divergence over horizon ticks. -1 = price up but flow down (bearish divergence), +1 = price down but flow up (bullish divergence), 0 = they agree.

tradeIntensityFeature - ticks per second over the last horizon ticks. How fast the tape's moving right now.

realizedVolatilityFeature - sqrt of summed squared tick-to-tick point returns over horizon bars. Realized vol without demeaning, unlike rollingVolatilityFeature.

returnSkewnessFeature - skewness of tick-to-tick point returns over horizon bars. Positive = occasional big up-moves, negative = same but down.

rollingDrawdownFeature - biggest peak-to-trough drop in points over horizon bars.

rangePositionFeature - where price sits in its own horizon-bar high/low range, rescaled to -1..1 (-1 at the low, +1 at the high).

autocorrelationFeature - lag-1 autocorrelation of tick-to-tick point returns over horizon bars. Positive = momentum regime, negative = mean-reverting/choppy regime.

efficiencyRatioFeature - Kaufman efficiency ratio over horizon bars, net move divided by total path length. Near 1 = clean trend, near 0 = noisy chop.

deltaMomentumFeature - acceleration of orderflow: CVD over the recent half of a horizon-tick window minus the prior half. Positive = flow speeding up.

spreadProxyFeature - average absolute tick-to-tick point move over horizon bars. Cheap noise/choppiness proxy when there's no real spread column.