io.deephaven.engine.table.impl.updateby.emstd.BigDecimalEmStdOperator Maven / Gradle / Ivy
Show all versions of deephaven-engine-table Show documentation
package io.deephaven.engine.table.impl.updateby.emstd;
import io.deephaven.api.updateby.OperationControl;
import io.deephaven.chunk.Chunk;
import io.deephaven.chunk.LongChunk;
import io.deephaven.chunk.attributes.Values;
import io.deephaven.engine.rowset.RowSequence;
import io.deephaven.engine.table.impl.MatchPair;
import io.deephaven.engine.table.impl.locations.TableDataException;
import io.deephaven.engine.table.impl.updateby.UpdateByOperator;
import org.jetbrains.annotations.NotNull;
import org.jetbrains.annotations.Nullable;
import java.math.BigDecimal;
import java.math.MathContext;
import static io.deephaven.util.QueryConstants.NULL_LONG;
/***
* Compute an exponential moving standard deviation for a BigDecimal column source. The output is expressed as a
* BigDecimal value and is computed using the following formula:
*
* variance = alpha * (prevVariance + (1 - alpha) * (x - prevEma)^2)
*
* This function is described in the following document:
*
* "Incremental calculation of weighted mean and variance" Tony Finch, University of Cambridge Computing Service
* (February 2009)
* https://web.archive.org/web/20181222175223/http://people.ds.cam.ac.uk/fanf2/hermes/doc/antiforgery/stats.pdf
*
* NOTE: `alpha` as used in the paper has been replaced with `1 - alpha` per the convention adopted by Deephaven.
*/
public class BigDecimalEmStdOperator extends BaseBigNumberEmStdOperator {
public class Context extends BaseBigNumberEmStdOperator.Context {
protected Context(final int affectedChunkSize, final int influencerChunkSize) {
super(affectedChunkSize);
}
@Override
public void accumulateCumulative(@NotNull final RowSequence inputKeys,
@NotNull final Chunk[] valueChunkArr,
@Nullable final LongChunk tsChunk,
final int len) {
setValueChunks(valueChunkArr);
// chunk processing
if (timestampColumnName == null) {
// compute with ticks
for (int ii = 0; ii < len; ii++) {
// read the value from the values chunk
final BigDecimal input = objectValueChunk.get(ii);
if (input == null) {
handleBadData(this, true);
} else {
if (curEma == null) {
curEma = input;
curVariance = BigDecimal.ZERO;
curVal = null;
} else {
// incremental variance = alpha * (prevVariance + (1 - alpha) * (x - prevEma)^2)
curVariance = opAlpha.multiply(
curVariance.add(
opOneMinusAlpha.multiply(input.subtract(curEma).pow(2, mathContext)),
mathContext),
mathContext);
final BigDecimal decayedEmaVal = curEma.multiply(opAlpha, mathContext);
curEma = decayedEmaVal.add(opOneMinusAlpha.multiply(input, mathContext));
curVal = curVariance.sqrt(mathContext);
}
}
outputValues.set(ii, curVal);
if (emaValues != null) {
emaValues.set(ii, curEma);
}
}
} else {
// compute with time
for (int ii = 0; ii < len; ii++) {
// read the value from the values chunk
final BigDecimal input = objectValueChunk.get(ii);
final long timestamp = tsChunk.get(ii);
final boolean isNull = input == null;
final boolean isNullTime = timestamp == NULL_LONG;
if (isNull) {
handleBadData(this, isNull);
} else if (isNullTime) {
// no change to curVal and lastStamp
} else if (curEma == null) {
// If the data looks good, and we have a null computed value, accept the current value
curEma = input;
lastStamp = timestamp;
} else {
final long dt = timestamp - lastStamp;
if (dt < 0) {
// negative time deltas are not allowed, throw an exception
throw new TableDataException("Timestamp values in UpdateBy operators must not decrease");
}
if (dt != 0) {
// alpha is dynamic based on time, but only recalculated when needed
if (dt != lastDt) {
alpha = computeAlpha(-dt, reverseWindowScaleUnits);
oneMinusAlpha = computeOneMinusAlpha(alpha);
lastDt = dt;
}
// incremental variance = alpha * (prevVariance + (1 - alpha) * (x - prevEma)^2)
curVariance = alpha.multiply(
curVariance.add(
oneMinusAlpha.multiply(input.subtract(curEma).pow(2, mathContext)),
mathContext),
mathContext);
final BigDecimal decayedEmaVal = curEma.multiply(alpha, mathContext);
curEma = decayedEmaVal.add(oneMinusAlpha.multiply(input, mathContext));
curVal = curVariance.sqrt(mathContext);
lastStamp = timestamp;
}
}
outputValues.set(ii, curVal);
if (emaValues != null) {
emaValues.set(ii, curEma);
}
}
}
// chunk output to column
writeToOutputColumn(inputKeys);
}
}
/**
* An operator that computes an EM Std from a BigDecimal column using an exponential decay function.
*
* @param pair the {@link MatchPair} that defines the input/output for this operation
* @param affectingColumns the names of the columns that affect this ema
* @param control defines how to handle {@code null} input values.
* @param timestampColumnName the name of the column containing timestamps for time-based calcuations
* @param windowScaleUnits the smoothing window for the EMA. If no {@code timestampColumnName} is provided, this is
* measured in ticks, otherwise it is measured in nanoseconds
*/
public BigDecimalEmStdOperator(@NotNull final MatchPair pair,
@NotNull final String[] affectingColumns,
@NotNull final OperationControl control,
@Nullable final String timestampColumnName,
final double windowScaleUnits,
@NotNull final MathContext mathContext) {
super(pair, affectingColumns, control, timestampColumnName, windowScaleUnits, mathContext);
}
@Override
public UpdateByOperator copy() {
return new BigDecimalEmStdOperator(
pair,
affectingColumns,
control,
timestampColumnName,
reverseWindowScaleUnits,
mathContext);
}
@NotNull
@Override
public UpdateByOperator.Context makeUpdateContext(final int affectedChunkSize, final int influencerChunkSize) {
return new Context(affectedChunkSize, influencerChunkSize);
}
}