Click to enlarge to full size.
The analogs selected for this TC (see left-hand side of this page for methodology) may be few or many, depending on how many TCs have passed near our storm's location in the past. We select up to the 10 best analogs, ranked by ACC value. The tracks of these storms are plotted in the top left of the figure (see example image above). Below the track map, the storms, dates, and ACC values are printed. It is important when using analogs to realize that you are NOT guaranteed to get any good analogs. It may be that a similar pattern at this time of year with a TC in this location has not occurred in the reanalysis period. Thus, note the ACC values. The analogs are loosely categorized as "good" if the ACC exceeds 0.6, "questionable" if the ACC is between 0.3 and 0.6, and "bad" if the ACC is less than 0.3. Good and questionable analogs usually have a similar-looking 500mb height pattern to the GFS 72-hour forecast, but not always. Qualitative analysis is key and necessary. This is where the right-hand side of the figure comes in.
On the right, a 4x4 subpanel plot is created. The top row shows the GFS 500mb height forecast at 72, 120, 168, and 240 hours (increasing left to right). The following 3 rows correspond to the 3 best analogs. The reanalysis time in each panel corresponds to the GFS forecast panel immediately above it in the top row. Because of how the analogs were chosen, it is expected that the analogs will be most similar to the GFS at the 72-hour forecast (left-most panels), though there are exceptions. Good analogs may mimic the GFS forecast all the way out to Day 10, though the GFS itself has substantial errors beyond Day 3, and thus the analogs are now just another form of predictive guidance, which is what they are intended to be. The location of the current TC in the GFS forecast or the analog TC in the reanalysis is indicated in each panel by a red dot if the TC exists in the model or the IBTrACS database at that time.
TC analogs will likely never exhibit greater track skill than today's dynamical numerical model guidance at short-range. Analogs are meant to provide insight into the medium to long-range forecast. This is especially useful in the tropics where today's global models still possess significant biases in the interaction between TCs and mid-latitude dynamical features (jet streaks, Rossby waves, cut-off lows, etc.), and these biases sour many TC forecasts beyond Day 5 or so. At lead times beyond ~5 days, however, the behavior and interactions of past TCs placed in similar locations and flow patterns may provide useful clues to the future evolution of a current TC and the pattern around it. If objectively good analogs exist for a storm, qualitative analysis can yield a likely position of the TC 5-10+ days into the future that may rival the skill of global models at that range.
The IBTrACS tropical cyclone database is used for historical storm data. IBTrACS is sanctioned by the World Meteorological Organization as an official global TC database. For the Atlantic, IBTrACS data is identical to HURDAT. Since IBTrACS uses the regional RSMC's data, storms outside of the NHC's jurisdiction are subject to the analysis procedures specific to that RSMC. For example, the western Pacific is handled by the JMA, which uses 10-min wind instead of 1-min wind, and does not have intensity information for TCs prior to the satellite era.
NCEP/NCAR Reanalysis I is used for historical meteorological data. While better-quality reanalysis sets exist (e.g. CFSR), their period of analysis is limited to the satellite era, and it is important when searching for analogs to use the largest possible sample of observed weather patterns. NCEP reanalysis is available from 1948 to the present (the analysis is always ongoing). The current period being used here is 1948-2013. As each year goes by, that period will grow to include the most recent years.
The search for analogs for an active TC begins by finding historical TCs which passed within 400 km of the current storm's position. Of these, we only keep storms which passed closest to the current TC during a 30-day window centered on the TC's analysis time. This is to limit seasonal differences between the present and historical patterns. We then track the current TC in the GFS model from 0-72 hours, or up until the storm dissipates or turns non-tropical in the model grid (see criteria here).
We now want to compare reanalysis Z500 with the 0-72hr Z500 forecast from the GFS. If we consider the GFS 0-72hr Z500 forecast to be "truth," correlating that with a reanalysis period corresponding to the forecast period will allow us to find analogs which "follow" the pattern well. In other words, unlike correlating with one timestep, correlating over a time interval allows new information like Rossby wave phase speed and dynamic flow evolution to get captured by an analog candidate. Of course, the analog pattern will still diverge from the current pattern as we project it into the future (beyond 72 hours), so we have to choose where we want the analog pattern to match best with the GFS. We will be correlating with the GFS 0, 24, 48, and 72-hour forecasts. We choose weights of [0.10, 0.20, 0.30, 0.40], linearly increasing with forecast time. This maximizes the analog's similarity to the GFS 72-hour forecast. Since a global model's 3-day forecast is very reliable, this makes 72 hours a great "launch point" for the analog. The analog will diverge more from the GFS analysis, but will converge toward the 72-hour forecast, after which it will begin diverging again further into the future.
We need the spatial domain for correlation to remain static so that the pattern evolution gets incorporated into the correlation. Thus, we choose the mean position of the current TC in the GFS 0-72hr forecast, which here we will call (TClat, TClon). The following domain is used: [TClon-40°, TClon+30°, TClat-15°, TClat+35°]. A greater area west and poleward of the TC is included in order to capture more influential features upstream from the TC. In the southern hemisphere, the meridional range would be [TClat-35°, TClat+15°]. In the north (south) hemisphere, if the poleward latitude boundary does not reach 60°N (60°S), it is extended to that value, in order to ensure that most of the jet stream and Rossby waves poleward of the TC are included in the domain.
For each historical TC selected earlier, we select all times at which the TC was within 400 km of the current TC. At each of these times, 500 hPa geopotential height from reanalysis is used to compute the anomaly correlation coefficient (ACC) with the corresponding GFS forecast. Thus, reanalysis time t is correlated with GFS analysis, t+24hr with GFS 24-hour forecast, and so on. The climatology used to calculate the anomalies used in the ACC is CFSR 1981-2010. The weighted average (using the weights described earlier) of these four ACC values is then taken to obtain a single ACC value for the analog time. Since we did this for all times at which the storm was within 400 km of the current TC, we take the time corresponding to the highest ACC score to be the analog time for that TC.
The analogs we have obtained can now be used in plotting. Look to the right underneath the example image for a description of what is plotted there.